Integrated system and method for data security protection

Abstract

The present application relates to the technical field of intelligent protection. Specifically disclosed in the present application are an integrated system and method for data security protection. In the present application, an artificial intelligence detection algorithm based on deep learning is used to extract feature information from data to be tested and a sensitive vocabulary, and a transition matrix between the features of the data to be tested and the features of the sensitive vocabulary is further calculated to represent a feature similarity between the data to be tested and the sensitive vocabulary, so as to determine whether the data is sensitive data. In this way, a large amount of data can be automatically processed, thereby improving the accuracy and efficiency of sensitive data identification.

Description

Data security protection integrated system and method thereof TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent protection, and more specifically, to a data security protection integrated system and method thereof. BACKGROUND

[0002] With the rapid development of the Internet, data security protection is attracting more and more attention, and sensitive data identification is of great significance in data security protection. The main purpose is to identify and mark sensitive data stored in the system, so as to take appropriate security measures to protect these data. Sensitive data refers to data that may cause serious harm to society or individuals if leaked. At the same time, sensitive data is also called private data, which includes all non-public or unclassified information, including personal private data such as name, ID number, address, phone number, bank account number, email, password, medical information, educational background, etc.; and enterprise private information such as enterprise operation, customer information, business secrets, etc.

[0003] Through sensitive data identification, these data can be discovered and marked in time, so as to take appropriate security measures to prevent unauthorized personnel from accessing, leaking or misusing them. Since the current data is mostly large in capacity and relatively complex, the traditional manual sorting speed is slow, and different people may have different judgments on the same data, resulting in differences in the results of sensitive data identification.

[0004] Therefore, a data security protection integrated system and method thereof are expected.

[0005] SUMMARY

[0006] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a data security protection integrated system and method thereof, which adopts an artificial intelligence detection algorithm based on deep learning to extract feature information of to-be-detected data and sensitive words, and further calculates a transfer matrix between the features of the to-be-detected data and the sensitive words to represent the feature similarity between them, so as to judge whether the data is sensitive data. In this way, a large amount of data can be processed automatically, and the accuracy and efficiency of sensitive data identification are improved.

[0007] Correspondingly, according to one aspect of the present application, a data security protection integrated system is provided, which comprises:

[0008] a data acquisition module, configured to acquire to-be-detected data and a sensitive word set;

[0009] a to-be-detected data semantic understanding module, configured to pass the to-be-detected data through a context encoder comprising an embedding layer to obtain a plurality of context semantic feature vectors;

[0010] a first scale perception module configured to concatenate the plurality of context semantic feature vectors to obtain a first scale semantic correlation feature vector;

[0011] a second scale perception module configured to arrange the plurality of context semantic feature vectors into a context semantic feature matrix, and then pass the context semantic feature matrix through a convolutional neural network model comprising a plurality of mixed convolutional layers to obtain a second scale semantic correlation feature vector;

[0012] a multi-scale fusion module configured to perform interpolation and ordering fusion on the first scale semantic correlation feature vector and the second scale semantic correlation feature vector to obtain a to-be-detected data feature vector;

[0013] a sensitive word semantic understanding module configured to pass the set of sensitive words through a context encoder comprising an embedding layer to obtain a sensitive data feature vector;

[0014] a transition calculation module configured to calculate a transition matrix between the to-be-detected data feature vector and the sensitive data feature vector as a classification feature matrix;

[0015] a detection result generation module configured to pass the classification feature matrix through a classifier to obtain a classification result, the classification result being used to indicate whether the detection data is sensitive data.

[0016] In the above-mentioned data security protection integrated system, the to-be-detected data semantic understanding module comprises: an embedding unit configured to use an embedding layer of the context encoder to respectively map text data of each data item in the to-be-detected data into a word embedding vector; a data adding unit configured to add numerical data in the each data item to a tail of the word embedding vector of the each data item to obtain a plurality of data item embedding vectors; and a context encoding unit configured to use a transformer-based Bert model of the context encoder to perform context semantic encoding on the plurality of data item embedding vectors to obtain the plurality of context semantic feature vectors.

[0017] In the data security protection integrated system, the context encoding unit comprises: a one-dimensional arrangement subunit configured to arrange the plurality of data item embedding vectors in one dimension to obtain a data item global embedding vector; a self-attention generation subunit configured to calculate the product between the data item global embedding vector and the transpose vector of each data item embedding vector in the plurality of data item embedding vectors to obtain a plurality of self-attention correlation matrices; a normalized self-attention subunit configured to respectively perform normalization processing on each self-attention correlation matrix in the plurality of self-attention correlation matrices to obtain a plurality of normalized self-attention correlation matrices; a weight generation subunit configured to pass each normalized self-attention correlation matrix in the plurality of normalized self-attention correlation matrices through a classification function to obtain a plurality of probability values; and a weighting subunit configured to respectively weight each data item embedding vector in the plurality of data item embedding vectors using each probability value in the plurality of probability values to obtain the plurality of context semantic feature vectors.

[0018] In the data security protection integrated system, the second scale perception module is configured to: use each mixed convolutional layer of the convolutional neural network model to respectively perform, in forward transmission of the layer, convolutional coding on the context semantic feature matrix using a first convolutional kernel with a first size to obtain a first scale feature map, convolutional coding on the context semantic feature matrix using a second convolutional kernel with a first hole rate to obtain a second scale feature map, convolutional coding on the context semantic feature matrix using a third convolutional kernel with a second hole rate to obtain a third scale feature map, and convolutional coding on the context semantic feature matrix using a fourth convolutional kernel with a third hole rate to obtain a fourth scale feature map, wherein the first convolutional kernel, the second convolutional kernel, the third convolutional kernel, and the fourth convolutional kernel have the same size, and the second convolutional kernel, the third convolutional kernel, and the fourth convolutional kernel have different hole rates; aggregate the first scale feature map, the second scale feature map, the third scale feature map, and the fourth scale feature map along the channel dimension to obtain an aggregated feature map; perform global pooling processing on each feature matrix of the aggregated feature map along the channel dimension to generate a pooled feature map; and perform activation processing on the pooled feature map to generate an activated feature map; wherein the output of the last layer of the convolutional neural network model comprising a plurality of mixed convolutional layers is the second scale semantic correlation feature vector.

[0019] In the aforementioned integrated data security protection system, the multi-scale fusion module includes: a difference calculation unit, used to calculate the positional difference between the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain a difference feature vector; a positional weighting unit, used to calculate the positional weighting between the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain a point-added feature vector; a cosine similarity calculation unit, used to calculate the cosine similarity between the difference feature vector and the point-added feature vector; and a weighted fusion unit, used to fuse the first-scale semantic association feature vector and the second-scale semantic association feature vector using the cosine similarity between the difference feature vector and the point-added feature vector as a weight parameter, and using the following fusion formula to obtain the feature vector of the data to be detected; wherein, the fusion formula is: V i = αV1 + (1-α)V2, where V1 represents the semantic association feature vector at the first scale, V2 represents the semantic association feature vector at the second scale, and α represents the weight parameter. i This represents the feature vector of the data to be detected.

[0020] In the aforementioned integrated data security protection system, the sensitive word semantic understanding module includes: an embedding vectorization unit, used to map each sensitive word in the sensitive word set into a word embedding vector using the embedding layer of the context encoder to obtain a sequence of word embedding vectors; a semantic encoding unit, used to perform global context semantic encoding on the sequence of word embedding vectors using the converter-based BERT model of the context encoder to obtain multiple word feature vectors; and a concatenation unit, used to concatenate the multiple word feature vectors to obtain the sensitive data feature vector.

[0021] In the aforementioned integrated data security protection system, the transfer calculation module is used to: calculate the transfer matrix between the feature vector of the data to be detected and the feature vector of the sensitive data using the following transfer formula;

[0022] The transfer formula is as follows:

[0023] Where V a V represents the feature vector of the data to be detected. b Let M represent the sensitive data feature vector, and let M represent the transition matrix. This represents the multiplication of a matrix and a vector.

[0024] According to another aspect of this application, an integrated data security protection method is provided, comprising:

[0025] Obtain the data to be detected and the set of sensitive words;

[0026] The data to be detected is passed through a context encoder containing an embedding layer to obtain multiple contextual semantic feature vectors;

[0027] The multiple context semantic feature vectors are concatenated to obtain a first-scale semantic association feature vector;

[0028] After arranging the multiple context semantic feature vectors in two dimensions into a context semantic feature matrix, the second-scale semantic association feature vector is obtained by using a convolutional neural network model containing multiple hybrid convolutional layers.

[0029] The semantic association feature vectors at the first scale and the semantic association feature vectors at the second scale are interpolated and ordered to be fused to obtain the feature vector of the data to be detected.

[0030] The sensitive vocabulary set is passed through a context encoder containing an embedding layer to obtain a sensitive data feature vector;

[0031] Calculate the transition matrix between the feature vector of the data to be detected and the feature vector of the sensitive data as the classification feature matrix;

[0032] The classification feature matrix is ​​passed through a classifier to obtain a classification result, which is used to indicate whether the detected data is sensitive data.

[0033] In the above-mentioned integrated data security protection method, the data to be detected is passed through a context encoder containing an embedding layer to obtain multiple context semantic feature vectors, including: using the embedding layer of the context encoder to map the text data of each data item in the data to be detected into word embedding vectors respectively; adding the numerical data in each data item to the end of the word embedding vector of each data item to obtain multiple data item embedding vectors; and using the converter-based BERT model of the context encoder to perform context semantic encoding on the multiple data item embedding vectors to obtain the multiple context semantic feature vectors.

[0034] In the aforementioned integrated data security protection method, the context encoder uses a transducer-based BERT model to perform contextual semantic encoding on the multiple data item embedding vectors to obtain the multiple contextual semantic feature vectors. This includes: arranging the multiple data item embedding vectors in a one-dimensional manner to obtain a global data item embedding vector; calculating the product between the global data item embedding vector and the transpose of each data item embedding vector in the multiple data item embedding vectors to obtain multiple self-attention association matrices; standardizing each self-attention association matrix in the multiple self-attention association matrices to obtain multiple standardized self-attention association matrices; passing each standardized self-attention association matrix in the multiple standardized self-attention association matrices through a classification function to obtain multiple probability values; and weighting each data item embedding vector in the multiple data item embedding vectors using each probability value as a weight to obtain the multiple contextual semantic feature vectors.

[0035] Compared with existing technologies, the integrated data security protection system and method provided in this application employs a deep learning-based artificial intelligence detection algorithm to extract feature information of the data to be detected and sensitive words. It then further calculates the transition matrix between the features of the data to be detected and the features of the sensitive words to represent their feature similarity, thereby determining whether the data is sensitive. This allows for automated processing of large amounts of data, improving the accuracy and efficiency of sensitive data identification. Attached Figure Description

[0036] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0037] Figure 1 is a block diagram of an integrated data security protection system according to an embodiment of this application.

[0038] Figure 2 is a schematic diagram of the architecture of an integrated data security protection system according to an embodiment of this application.

[0039] Figure 3 is a block diagram of the semantic understanding module of the data to be detected in the integrated data security protection system according to an embodiment of this application.

[0040] Figure 4 is a block diagram of the context coding unit in the integrated data security protection system according to an embodiment of this application.

[0041] Figure 5 is a block diagram of the sensitive word semantic understanding module in the integrated data security protection system according to an embodiment of this application.

[0042] Figure 6 is a flowchart of an integrated data security protection method according to an embodiment of this application. Detailed Implementation

[0043] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0044] Figure 1 is a block diagram of an integrated data security protection system according to an embodiment of this application. As shown in Figure 1, the integrated data security protection system 100 according to an embodiment of this application includes: a data acquisition module 110, used to acquire data to be detected and a set of sensitive words; a semantic understanding module 120 for processing the data to be detected through a context encoder including an embedding layer to obtain multiple context semantic feature vectors; a first scale perception module 130 for concatenating the multiple context semantic feature vectors to obtain a first scale semantic association feature vector; and a second scale perception module 140 for arranging the multiple context semantic feature vectors in a two-dimensional manner into a context semantic feature matrix and then processing it through a convolutional neural network model including multiple hybrid convolutional layers to obtain a second scale semantic association feature vector. The system comprises: a semantic association feature vector; a multi-scale fusion module 150, used to interpolate and order-fuse the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain a feature vector of the data to be detected; a sensitive word semantic understanding module 160, used to pass the sensitive word set through a context encoder containing an embedding layer to obtain a sensitive data feature vector; a transition calculation module 170, used to calculate the transition matrix between the feature vector of the data to be detected and the sensitive data feature vector as a classification feature matrix; and a detection result generation module 180, used to pass the classification feature matrix through a classifier to obtain a classification result, the classification result being used to indicate whether the detected data is sensitive data.

[0045] Figure 2 is a schematic diagram of the architecture of an integrated data security protection system according to an embodiment of this application. As shown in Figure 2, firstly, the data to be detected and a set of sensitive words are acquired. Then, the data to be detected is passed through a context encoder containing an embedding layer to obtain multiple context semantic feature vectors. Next, the multiple context semantic feature vectors are concatenated to obtain a first-scale semantic association feature vector. Simultaneously, the multiple context semantic feature vectors are arranged in two dimensions into a context semantic feature matrix and then passed through a convolutional neural network model containing multiple hybrid convolutional layers to obtain a second-scale semantic association feature vector. Subsequently, the first-scale semantic association feature vector and the second-scale semantic association feature vector are interpolated and ordered to obtain a feature vector of the data to be detected. Secondly, the set of sensitive words is passed through a context encoder containing an embedding layer to obtain a sensitive data feature vector. Next, the transition matrix between the feature vector of the data to be detected and the sensitive data feature vector is calculated as a classification feature matrix. Finally, the classification feature matrix is ​​passed through a classifier to obtain a classification result, which is used to indicate whether the detected data is sensitive data.

[0046] In the aforementioned integrated data security protection system 100, the data acquisition module 110 is used to acquire the data to be detected and a set of sensitive words. As mentioned in the background section, sensitive data identification is of great significance in data security protection. Its main purpose is to identify and mark sensitive data stored in the system so that corresponding security measures can be taken to prevent it from being accessed, leaked, or misused by unauthorized personnel. However, since most data today is characterized by its large volume and complexity, and traditional manual sorting is slow, and different people may have different judgments about the same data, the results of sensitive data identification will vary. Therefore, an efficient and accurate sensitive data identification scheme is desired.

[0047] Accordingly, in order to balance accuracy and effectiveness when identifying sensitive data, the feature similarity between the implicit features of the data to be detected and the implicit features of sensitive words in a high-dimensional space can be used. That is, in the technical solution of this application, a deep learning-based artificial intelligence detection algorithm is used to extract the feature information of the data to be detected and the sensitive words, and then the transition matrix between the features of the data to be detected and the features of the sensitive words is calculated to represent the feature similarity between the two, thereby determining whether the data is sensitive data. This allows for automated processing of large amounts of data, improving the accuracy and efficiency of sensitive data identification. Specifically, in the technical solution of this application, firstly, a set of data to be detected and a set of sensitive words are obtained.

[0048] In the aforementioned integrated data security protection system 100, the semantic understanding module 120 for the data to be detected is used to process the data to be detected through a context encoder containing an embedding layer to obtain multiple context semantic feature vectors. In sensitive data identification, the contextual information of the data is crucial for accurately determining whether the data is sensitive. Therefore, in order to capture the contextual information and semantic association features in the data to be detected, the data to be detected is processed through a context encoder containing an embedding layer. It should be understood that a context encoder is a model used to convert text data into a continuous vector representation. First, through the embedding layer, each word or character in the text data is mapped to an embedding vector representation; then, by using the converter of the context encoder, a self-attention mechanism is introduced into multiple embedding vectors to convert multiple embedding vectors into multiple context semantic feature vectors, comprehensively capturing the semantic association features of the data, thereby improving the accuracy of sensitive data identification.

[0049] Figure 3 is a block diagram of the semantic understanding module of the data to be detected in the integrated data security protection system according to an embodiment of this application. As shown in Figure 3, the semantic understanding module 120 of the data to be detected includes: an embedding unit 121, used to map the text data of each data item in the data to be detected into word embedding vectors using the embedding layer of the context encoder; a data adding unit 122, used to add the numerical data in each data item to the end of the word embedding vector of each data item to obtain multiple data item embedding vectors; and a context encoding unit 123, used to perform context semantic encoding on the multiple data item embedding vectors using the converter-based BERT model of the context encoder to obtain the multiple context semantic feature vectors.

[0050] Figure 4 is a block diagram of the context encoding unit in the integrated data security protection system according to an embodiment of this application. The context encoding unit 123 includes: a one-dimensional arrangement subunit 1231, used to arrange the plurality of data item embedding vectors in one dimension to obtain a global data item embedding vector; a self-attention generation subunit 1232, used to calculate the product between the global data item embedding vector and the transpose of each data item embedding vector in the plurality of data item embedding vectors to obtain a plurality of self-attention association matrices; a standardized self-attention subunit 1233, used to standardize each self-attention association matrix in the plurality of self-attention association matrices to obtain a plurality of standardized self-attention association matrices; a weight generation subunit 1234, used to pass each standardized self-attention association matrix in the plurality of standardized self-attention association matrices through a classification function to obtain a plurality of probability values; and a weighting subunit 1235, used to weight each data item embedding vector in the plurality of data item embedding vectors with each probability value in the plurality of probability values ​​as weights to obtain a plurality of context semantic feature vectors.

[0051] In the aforementioned integrated data security protection system 100, the first scale perception module 130 is used to concatenate the multiple context semantic feature vectors to obtain a first-scale semantic association feature vector. To comprehensively consider the semantic information in the multiple context semantic feature vectors and capture the more global semantic association features of the data to be detected, it is necessary to fuse the multiple context semantic feature vectors. Here, a concatenation operation is used to integrate the multiple context semantic feature vectors to form a more global feature representation, thereby better reflecting the overall semantic association features of the data to be detected.

[0052] In the aforementioned integrated data security protection system 100, the second scale perception module 140 is used to arrange the multiple context semantic feature vectors in a two-dimensional manner into a context semantic feature matrix, and then pass it through a convolutional neural network model containing multiple hybrid convolutional layers to obtain a second-scale semantic association feature vector. To further extract and capture the local and global semantic association features of the data, a convolutional data network model is used to process the multiple context semantic feature vectors. It should be understood that arranging the multiple context semantic feature vectors into the form of the context semantic feature matrix allows for the organization and representation of different context semantic information in a two-dimensional space, helping the model to better understand the relationships between contexts and extract local semantic association features. Simultaneously, convolutional neural networks have excellent feature extraction capabilities in image and text processing, while hybrid convolutional layers can simultaneously capture the association between local details and global context. Therefore, by applying convolutional operations to the context semantic feature matrix, semantic association features at different scales can be effectively extracted from the context semantic feature matrix.

[0053] Accordingly, in a specific example, the second scale-aware module 140 is configured to: use the various hybrid convolutional layers of the convolutional neural network model to perform the following operations on the input data during the forward propagation of the layers: convolutionally encode the context semantic feature matrix using a first convolutional kernel with a first size to obtain a first-scale feature map; convolutionally encode the context semantic feature matrix using a second convolutional kernel with a first dilation rate to obtain a second-scale feature map; convolutionally encode the context semantic feature matrix using a third convolutional kernel with a second dilation rate to obtain a third-scale feature map; and convolutionally encode the context semantic feature matrix using a fourth convolutional kernel with a third dilation rate to obtain a fourth-scale feature map. The first, second, third, and fourth convolutional kernels have the same size, and the second, third, and fourth convolutional kernels have different dilatation rates. The first, second, third, and fourth scale feature maps are aggregated along the channel dimension to obtain an aggregated feature map. Global pooling is performed on each feature matrix along the channel dimension of the aggregated feature map to generate a pooled feature map. Activation processing is performed on the pooled feature map to generate an activated feature map. The output of the last layer of the convolutional neural network model containing multiple hybrid convolutional layers is the second-scale semantic association feature vector.

[0054] In the aforementioned integrated data security protection system 100, the multi-scale fusion module 150 is used to perform interpolation-ordered fusion of the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain the feature vector of the data to be detected. To comprehensively utilize the semantic association features at different scales, the first-scale semantic association feature vector and the second-scale semantic association feature vector are further fused. It should be understood that the first-scale semantic association feature vector focuses more on the overall semantic association features of the data, while the second-scale semantic association feature vector focuses more on local and global semantic associations. By fusing the features of these two scales, the different characteristics and information of both can be comprehensively utilized, thereby more comprehensively describing and representing the data to be detected, and improving the accuracy and robustness of sensitive data identification.

[0055] Specifically, in the technical solution of this application, the first-scale semantic association feature vector and the second-scale semantic association feature vector represent different high-dimensional feature manifolds in the high-dimensional feature space, but in the class probability label domain, the first-scale semantic association feature vector and the second-scale semantic association feature vector point to the same class probability label. Therefore, the high-dimensional feature manifolds of the first-scale semantic association feature vector and the second-scale semantic association feature vector have an implicit correlation at the manifold representation level. That is to say, in the technical solution of this application, the high-dimensional data manifolds of the first-scale semantic association feature vector and the second-scale semantic association feature vector have smoothness and robustness at the manifold representation level. Based on this, in the technical solution of this application, firstly, positional difference and positional addition are used to represent the manifold difference and manifold superposition extension of the first-scale semantic association feature vector and the second-scale semantic association feature vector in the high-dimensional feature space. Then, the cosine similarity between the difference feature vector and the added feature vector is used to represent the smoothness and robustness of the high-dimensional data manifolds of the first-scale semantic association feature vector and the second-scale semantic association feature vector at the manifold representation level. Furthermore, using the cosine similarity between the difference feature vector and the point-plus feature vector as a weight parameter, the first-scale semantic association feature vector and the second-scale semantic association feature vector are fused using the following formula to obtain the feature vector of the data to be detected, wherein the formula is: V i = αV1 + (1-α)V2, where V1 represents the semantic association feature vector at the first scale, V2 represents the semantic association feature vector at the second scale, and α represents the weight parameter. i Let represent the feature vector of the data to be detected. Thus, the high-dimensional feature manifold of the feature vector of the data to be detected is geometrically collinear with the first-scale semantic association feature vector and the second-scale semantic association feature vector, but the manifold paradigm and manifold metric differ. Algebraically, there is manifold transformation consistency. Therefore, the feature vector of the data to be detected can utilize the high-dimensional implicit correlation between the first-scale semantic association feature vector and the second-scale semantic association feature vector for feature fusion to improve the smoothness and robustness of the fused feature vector of the data to be detected, thereby improving the accuracy of the final classification feature matrix in the classification judgment of the classifier.

[0056] Accordingly, in a specific example, the multi-scale fusion module 150 includes: a difference calculation unit, used to calculate the positional difference between the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain a difference feature vector; a positional weighting unit, used to calculate the positional weighting between the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain a point-added feature vector; a cosine similarity calculation unit, used to calculate the cosine similarity between the difference feature vector and the point-added feature vector; and a weighted fusion unit, used to fuse the first-scale semantic association feature vector and the second-scale semantic association feature vector using the cosine similarity between the difference feature vector and the point-added feature vector as a weight parameter, and using the following fusion formula to obtain the feature vector of the data to be detected; wherein, the fusion formula is: V i = αV1 + (1-α)V2, where V1 represents the semantic association feature vector at the first scale, V2 represents the semantic association feature vector at the second scale, and α represents the weight parameter. i This represents the feature vector of the data to be detected.

[0057] In the aforementioned integrated data security protection system 100, the sensitive word semantic understanding module 160 is used to obtain sensitive data feature vectors by passing the sensitive word set through a context encoder containing an embedding layer. Considering that the sensitive word set is usually given in text form, however, it is inconvenient to directly compare and calculate sensitive words with the data to be detected. Therefore, it is necessary to use a context encoder to mine the contextual semantic features of sensitive words and transform the sensitive word set into a computable and comparable vector representation so as to compare and classify features with the data to be detected.

[0058] Figure 5 is a block diagram of the sensitive word semantic understanding module in the integrated data security protection system according to an embodiment of this application. As shown in Figure 5, the sensitive word semantic understanding module 160 includes: an embedding vectorization unit 161, used to map each sensitive word in the sensitive word set into a word embedding vector using the embedding layer of the context encoder to obtain a sequence of word embedding vectors; a semantic encoding unit 162, used to perform global context semantic encoding on the sequence of word embedding vectors using the converter-based BERT model of the context encoder to obtain multiple word feature vectors; and a concatenation unit 163, used to concatenate the multiple word feature vectors to obtain the sensitive data feature vector.

[0059] In the aforementioned integrated data security protection system 100, the transfer calculation module 170 is used to calculate the transfer matrix between the feature vector of the data to be detected and the feature vector of the sensitive data as a classification feature matrix. To capture the semantic association and transfer features between the data to be detected and the sensitive data, the transfer matrix between the feature vector of the data to be detected and the feature vector of the sensitive data is further calculated. It should be understood that the transfer matrix can be viewed as a transformation matrix mapping the feature vector of the data to be detected to the feature vector of the sensitive data. Each element in the matrix represents the transfer relationship between a certain dimension of the feature vector of the data to be detected and the corresponding dimension of the feature vector of the sensitive data, reflecting the similarity, difference, and degree of semantic transfer between the data to be detected and the sensitive data. Based on this feature information, a classifier is then used to determine whether the data to be detected belongs to the sensitive data category.

[0060] Accordingly, in a specific example, the transfer calculation module 170 is used to: calculate the transfer matrix between the feature vector of the data to be detected and the feature vector of the sensitive data using the following transfer formula;

[0061] The transfer formula is as follows:

[0062] Where V a V represents the feature vector of the data to be detected. b Let M represent the sensitive data feature vector, and let M represent the transition matrix. This represents the multiplication of a matrix and a vector.

[0063] In the aforementioned integrated data security protection system 100, the detection result generation module 180 is used to pass the classification feature matrix through a classifier to obtain a classification result, which indicates whether the detected data is sensitive data. A classifier is a trained machine learning model. The training process of a classifier is usually based on labeled training data, which includes the features of the data to be detected and the corresponding category labels (sensitive data or non-sensitive data). Through training, the classifier can learn the association between features and categories and complete the classification of unknown data. Here, using the classification feature matrix as input, a classification result indicating whether the data to be detected is sensitive data can be obtained. This achieves automated classification of the data to be detected and identification of sensitive data, and further processing or decision-making based on the classification result, thereby ensuring data security.

[0064] In summary, the integrated data security protection system according to the embodiments of this application is explained. It employs a deep learning-based artificial intelligence detection algorithm to extract feature information of the data to be detected and sensitive words. Furthermore, it calculates the transition matrix between the features of the data to be detected and the features of the sensitive words to represent their feature similarity, thereby determining whether the data is sensitive. This allows for automated processing of large amounts of data, improving the accuracy and efficiency of sensitive data identification.

[0065] Figure 6 is a flowchart of an integrated data security protection method according to an embodiment of this application. As shown in Figure 6, the integrated data security protection method according to an embodiment of this application includes the following steps: S110, acquiring the data to be detected and a set of sensitive words; S120, passing the data to be detected through a context encoder including an embedding layer to obtain multiple context semantic feature vectors; S130, concatenating the multiple context semantic feature vectors to obtain a first-scale semantic association feature vector; S140, arranging the multiple context semantic feature vectors in a two-dimensional manner into a context semantic feature matrix, and then passing it through a convolutional neural network model including multiple hybrid convolutional layers to obtain a second-scale semantic association feature vector; S150, performing interpolation and ordering fusion on the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain a feature vector of the data to be detected; S160, passing the set of sensitive words through a context encoder including an embedding layer to obtain a sensitive data feature vector; S170, calculating the transition matrix between the feature vector of the data to be detected and the sensitive data feature vector as a classification feature matrix; S180, passing the classification feature matrix through a classifier to obtain a classification result, the classification result being used to indicate whether the detected data is sensitive data.

[0066] Here, those skilled in the art will understand that the specific operations of each step in the above-described integrated data security protection method have been described in detail in the description of the integrated data security protection system with reference to Figures 1 to 5, and therefore, their repeated description will be omitted.

Claims

1. An integrated data security protection system, characterized in that, include: The data acquisition module is used to acquire the data to be detected and the set of sensitive words; The data to be detected semantic understanding module is used to pass the data to be detected through a context encoder containing an embedding layer to obtain multiple context semantic feature vectors; The first scale perception module is used to concatenate the multiple context semantic feature vectors to obtain a first scale semantic association feature vector. The second-scale perception module is used to arrange the multiple context semantic feature vectors in two dimensions into a context semantic feature matrix, and then use a convolutional neural network model containing multiple hybrid convolutional layers to obtain the second-scale semantic association feature vector. The multi-scale fusion module is used to perform interpolation and ordering fusion on the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain the feature vector of the data to be detected. A sensitive word semantic understanding module is used to pass the set of sensitive words through a context encoder containing an embedding layer to obtain a sensitive data feature vector; The transition calculation module is used to calculate the transition matrix between the feature vector of the data to be detected and the feature vector of the sensitive data as a classification feature matrix. The detection result generation module is used to pass the classification feature matrix through a classifier to obtain a classification result, which is used to indicate whether the detection data is sensitive data.

2. The integrated data security protection system according to claim 1, characterized in that, The semantic understanding module for the data to be detected includes: An embedding unit is used to map the text data of each data item in the data to be detected into word embedding vectors using the embedding layer of the context encoder. A data addition unit is used to add the numerical data in each data item to the end of the word embedding vector of each data item to obtain multiple data item embedding vectors. The context encoding unit is used to perform context semantic encoding on the multiple data item embedding vectors using the converter-based BERT model of the context encoder to obtain the multiple context semantic feature vectors.

3. The integrated data security protection system according to claim 2, characterized in that, The context coding unit includes: A one-dimensional arrangement subunit is used to arrange the multiple data item embedding vectors in one dimension to obtain a global data item embedding vector. The self-attention generation subunit is used to calculate the product between the global embedding vector of the data item and the transpose of the embedding vector of each data item in the plurality of data item embedding vectors to obtain a plurality of self-attention association matrices. A standardized self-attention subunit is used to standardize each of the multiple self-attention association matrices to obtain multiple standardized self-attention association matrices. The weight generation subunit is used to obtain multiple probability values ​​by passing each of the multiple standardized self-attention association matrices through a classification function. The weighting subunit is used to weight each data item embedding vector in the plurality of data item embedding vectors by using each probability value among the plurality of probability values ​​as weights to obtain the plurality of context semantic feature vectors.

4. The integrated data security protection system according to claim 3, characterized in that, The second scale-aware module is used to: process the input data using each hybrid convolutional layer of the convolutional neural network model during the forward propagation of the layer: The context semantic feature matrix is ​​convolutionally encoded using a first convolution kernel of a first size to obtain a first-scale feature map; The context semantic feature matrix is ​​convolutionally encoded using a second convolution kernel with a first dilation rate to obtain a second-scale feature map. The context semantic feature matrix is ​​convolutionally encoded using a third convolution kernel with a second dilation rate to obtain a third-scale feature map; The context semantic feature matrix is ​​convolutionally encoded using a fourth convolution kernel with a third dilation rate to obtain a fourth-scale feature map, wherein the first, second, third, and fourth convolution kernels have the same size, and the second, third, and fourth convolution kernels have different dilation rates. The first-scale feature map, the second-scale feature map, the third-scale feature map, and the fourth-scale feature map are aggregated along the channel dimension to obtain an aggregated feature map; The aggregated feature map is subjected to global pooling along the channel dimension of each feature matrix to generate a pooled feature map. The pooled feature map is activated to generate an activated feature map; The output of the last layer of the convolutional neural network model containing multiple hybrid convolutional layers is the second-scale semantic association feature vector.

5. The integrated data security protection system according to claim 4, characterized in that, The multi-scale fusion module includes: The difference calculation unit is used to calculate the positional difference between the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain the difference feature vector; The position-weighted unit is used to calculate the position-weighted feature vector between the first-scale semantic association feature vector and the second-scale semantic association feature vector to obtain the point-weighted feature vector; The cosine similarity calculation unit is used to calculate the cosine similarity between the difference feature vector and the point-plus feature vector; The weighted fusion unit is used to fuse the first-scale semantic association feature vector and the second-scale semantic association feature vector using the cosine similarity between the difference feature vector and the point-addition feature vector as weight parameters, and to obtain the feature vector of the data to be detected using the following fusion formula; wherein, the fusion formula is: V i = αV1 + (1-α)V2, where V1 represents the semantic association feature vector at the first scale, V2 represents the semantic association feature vector at the second scale, and α represents the weight parameter. i This represents the feature vector of the data to be detected.

6. The integrated data security protection system according to claim 5, characterized in that, The sensitive word semantic understanding module includes: An embedding vectorization unit is used to map each sensitive word in the sensitive word set into a word embedding vector using the embedding layer of the context encoder to obtain a sequence of word embedding vectors; A semantic encoding unit is used to perform global contextual semantic encoding on the sequence of word embedding vectors using the converter-based BERT model of the context encoder to obtain multiple word feature vectors. A cascade unit is used to cascade the multiple word feature vectors to obtain the sensitive data feature vector.

7. The integrated data security protection system according to claim 6, characterized in that, The transfer calculation module is used to: calculate the transfer matrix between the feature vector of the data to be detected and the feature vector of the sensitive data using the following transfer formula; The transfer formula is as follows: Where V a V represents the feature vector of the data to be detected. b Let M represent the sensitive data feature vector, and let M represent the transition matrix. This represents the multiplication of a matrix and a vector.

8. An integrated data security protection method, characterized in that, include: Obtain the data to be detected and the set of sensitive words; The data to be detected is passed through a context encoder containing an embedding layer to obtain multiple contextual semantic feature vectors; The multiple context semantic feature vectors are concatenated to obtain a first-scale semantic association feature vector; After arranging the multiple context semantic feature vectors in two dimensions into a context semantic feature matrix, the second-scale semantic association feature vector is obtained by using a convolutional neural network model containing multiple hybrid convolutional layers. The semantic association feature vectors at the first scale and the semantic association feature vectors at the second scale are interpolated and ordered to be fused to obtain the feature vector of the data to be detected. The sensitive vocabulary set is passed through a context encoder containing an embedding layer to obtain a sensitive data feature vector; Calculate the transition matrix between the feature vector of the data to be detected and the feature vector of the sensitive data as the classification feature matrix; The classification feature matrix is ​​passed through a classifier to obtain a classification result, which is used to indicate whether the detected data is sensitive data.

9. The integrated data security protection method according to claim 8, characterized in that, The data to be detected is passed through a context encoder containing an embedding layer to obtain multiple contextual semantic feature vectors, including: The embedding layer of the context encoder is used to map the text data of each data item in the data to be detected into word embedding vectors respectively; The numerical data in each data item is added to the end of the word embedding vector of each data item to obtain multiple data item embedding vectors. The context semantics of the multiple data item embedding vectors are encoded using the converter-based BERT model of the context encoder to obtain the multiple context semantic feature vectors.

10. The integrated data security protection method according to claim 9, characterized in that, The context encoder uses a converter-based BERT model to perform context semantic encoding on the embedding vectors of the plurality of data items to obtain the plurality of context semantic feature vectors, including: The multiple data item embedding vectors are arranged in one dimension to obtain the global data item embedding vector; Calculate the product between the global embedding vector of the data item and the transpose of the embedding vector of each data item in the plurality of data item embedding vectors to obtain a plurality of self-attention association matrices; Standardize each of the multiple self-attention correlation matrices. The process was optimized to obtain multiple standardized self-attention correlation matrices; Each of the multiple standardized self-attention association matrices is classified using a classification function to obtain multiple probability values. The embedding vectors of each data item are weighted by using each of the multiple probability values ​​as weights to obtain the multiple context semantic feature vectors.

Images