TWI769531B - Document confidentiality level management system and method - Google Patents

Abstract

A file confidentiality level management system, comprising: a plurality of file confidentiality level setting systems and a federated learning system, each file confidentiality level setting system analyzes the file confidentiality level according to an analysis model, and provides a first confidentiality level according to the analysis result The label is compared with a second confidentiality level label provided by the user. If they are the same, the document is encrypted or authorized according to the first confidentiality level label; if different, the document is encrypted or Authority control; each file confidentiality level setting system encrypts and transmits the model parameters of each analysis model to the federated learning system; the federated learning system generates an update parameter and transmits it to each file confidentiality level setting system, so as to update the analysis model with the updated parameters. This improves accuracy and confidentiality.

Description

Document confidentiality level management system and method

The present invention is related to the technology for setting the confidentiality level of data files; in particular, it refers to a system and method for setting the confidentiality level of documents using artificial intelligence.

Traditionally, most of the document's confidentiality level is set manually, and the document's confidentiality level is determined based on manual review of specific information in the document, so as to determine the document's confidentiality level.

With the development of science and technology, documents written on paper in the past have been gradually presented in the form of electronic documents. Unlike paper documents, which are not easy to modify, increase or decrease content, electronic documents can directly use electronic software to modify the content of the document. Among them, the current technology for setting the confidentiality level of a document is only to judge the confidentiality level of a document through contents such as filenames or keywords, or to classify the confidentiality level of a document simply by manually reading the document.

In addition, when the content of a file is increased or decreased or re-edited, its confidentiality level may change. Especially now that the content of electronic files is edited and modified very quickly, it is necessary to conduct an immediate analysis of the confidentiality level of the file, or even re-judgment the file in real time. The confidentiality level of the classification personnel is a big burden.

In addition, when classifying personnel to judge the confidentiality level of documents, because they are used to the frequent words used in companies or units, when rare or new words involving confidentiality appear in the documents, the classification personnel are not easy to detect. It may have misjudged the confidentiality level of the document.

In view of this, the purpose of the present invention is to provide a document confidentiality level management system and method, which can classify the document's confidentiality level according to artificial intelligence after the document is generated or edited, and add the user's judgment of the document's confidentiality level , to decide which confidentiality level to use to control the document, and at the same time, it can optimize the accuracy of the classification of the document's confidentiality level by artificial intelligence.

In order to achieve the above object, the present invention provides a document confidentiality level management system, which includes: a plurality of document confidentiality level setting systems, each of which includes an artificial intelligence analysis module and a data leakage protection server. , wherein: the artificial intelligence analysis module is used for receiving a file, and analyzes the confidentiality level of the file according to an analysis model, and provides a first confidentiality level label according to the analysis result; the data leakage prevention server receives the The first security level label is connected with a terminal device, the terminal device is used for user operation and provides a second security level label, and the data leakage prevention server compares the first security level label with the first security level label. Whether the two confidentiality labels are the same, if they are the same, the data leakage prevention server marks the file with the first confidentiality label, and encrypts or controls the file according to the content defined by the first confidentiality label. management; if not, the data leakage prevention server marks the file with the second-class label, and encrypts or controls the file according to the content defined by the second-class label; a federated learning The system is connected with each of the document confidentiality level setting systems, and each of the document confidentiality level setting systems encrypts a model parameter of each of the analysis models to form an encrypted parameter and transmits it to the federated learning system; the federated learning system aggregates these encrypted parameters And generating an update parameter, and sending the update parameter to each of the document confidentiality level setting systems, and each of the artificial intelligence analysis modules to update the analysis model according to the update parameter.

In order to achieve the above object, the present invention further provides a document confidentiality level management method, which is applied in the document confidentiality level management system, and the method includes the following steps:

Each of the document confidentiality level setting systems performs the following steps:

A1. A terminal device provides a file;

A2. The document is analyzed by the artificial intelligence analysis module according to an analysis model for the confidentiality level of the document, and the artificial intelligence analysis module provides a first confidentiality level label to the data leakage prevention server according to the analysis result;

A3. The document is analyzed by a user for the confidentiality level, and a second confidentiality level label is provided to the data leakage prevention server;

A4. The data leakage prevention server compares the similarities and differences between the first confidentiality level label and the second confidentiality level label; when the first confidentiality level label is the same as the second confidentiality level label, the data leakage prevention server The device marks the file with the first confidentiality level label, and encrypts or controls the file according to the content defined by the first confidentiality level label; when the first confidentiality level label is different from the second confidentiality level label When the data leakage prevention server marks the document with the second confidentiality level label, the data leakage prevention server encrypts or controls the authority of the document according to the content defined by the second confidentiality level label;

A5. Encrypt a model parameter of the analysis model to form an encrypted parameter and transmit it to the federated learning system;

The federated learning system performs the following steps:

B1. Receive the encryption parameters from the document confidentiality level setting system;

B2. Aggregate these encryption parameters and generate an update parameter;

B3. Transmit the update parameter to each of the document confidentiality level setting systems;

Each of the document confidentiality level setting system also performs the following steps:

A6. Each of the document confidentiality level setting systems receives the update parameter;

A7. Each of the artificial intelligence analysis modules updates the analysis model according to the update parameter.

The effect of the present invention is that each of the document confidentiality level setting systems uses the artificial intelligence analysis module to analyze the confidentiality level of the document, and uses the user to analyze the confidentiality level of the same document, and compares the artificial intelligence analysis module with the same document. Whether there is a discrepancy in the confidentiality level given by the user, and control the authority of the document. And each of the document confidentiality level setting systems further encrypts the model parameters of the analysis model and transmits them to the federated learning system to obtain updated parameters, and each of the document confidentiality level setting systems updates the original analysis model according to the updated parameters to improve artificial intelligence analysis. Accuracy of module predictions. Since the model parameters of each document confidentiality level setting system are encrypted and sent out, instead of directly transmitting the document or analysis model, it can effectively avoid the leakage of confidential documents and improve the document confidentiality level setting system for each document. data security.

In addition, the present invention can be applied to the situation where the training data volume of each document confidentiality level setting system is insufficient. By using the parameters of other document confidentiality level setting systems, the problem of insufficient training data volume of each document security level setting system is effectively solved, and the accuracy is improved. sex.

Let multiple users with similar confidentiality level identification standards use the same document confidentiality level setting system, so as to avoid the use of a single document confidentiality level setting system for different users who have too large a difference in the confidentiality level identification, resulting in inaccurate analysis models.

In order to illustrate the present invention more clearly, the following examples are given and detailed descriptions are given in conjunction with the drawings. Please refer to FIG. 1 . FIG. 1 is a block diagram of a system for setting a document confidentiality level according to an embodiment of the present invention. The document confidentiality level setting system includes: a data leakage prevention server 10 and an artificial intelligence server 20 . The artificial intelligence server 20 includes an artificial intelligence analysis module 30 and an artificial intelligence training module 40 .

The data leakage protection server 10 is connected to at least one terminal device 1, and the data leakage protection server 10 manages the files processed by the terminal device 1 according to a data leakage protection specification, for example, the data leakage protection The specification records that when a specific file or file format is created, opened, shared, accessed, edited, deleted, or other file processing operations, the data leakage prevention server 10 will be based on the terminal device 1 or the The authorized scope of the user of the terminal device 1 determines whether the file processing action can be executed. In this embodiment, the data leakage protection server 10 further includes a data leakage protection unit 12, and the data leakage protection unit 12 may be, but is not limited to, an application for installation on the terminal device 1, It is connected to the data leakage prevention server 10, and can upload the file to the data leakage prevention server 10, and then the data leakage prevention server 10 transmits the file to the artificial intelligence of the artificial intelligence server 20. The analysis module 30 is used to analyze the confidentiality level of the document. The data leakage protection unit 12 manages the files processed by the terminal device 1 according to the data leakage protection specification. More specifically, the data leakage protection unit 12 can monitor the files processed by the user through the terminal device 1 . For example, in the data leakage protection specification, it can be set that when the file opened by the terminal device 1 is a file type such as a word file, a PDF file, or a file of other specific file types, or the content of the file edited by the terminal device 1 When sensitive words or sentences are involved in the file, the data leakage prevention unit 12 will activate and monitor the file processing action of the terminal device 1 for the file, and according to the content corresponding to the confidentiality level label marked on the file, Encrypt or control the authority of the file, wherein the encryption may be but not limited to hiding or obscuring or other encryption methods to make specific or sensitive words or sentences in the file, so that they cannot be used normally on the terminal device. Displayed or cannot be read by the user; the permission control may be, but not limited to, restricting the user's operation behavior or instructions on the file on the terminal device 1, such as including but not limited to restricting the file or The content of the file performs file processing operations such as opening, sharing, accessing, editing, deleting, copying, and clipping. In actual operation, there may be a plurality of the data leakage prevention units 12, which are respectively installed on different terminal devices 1 to monitor each terminal device 1 respectively. The terminal devices 1 may be, but are not limited to, mobile phones. , Tablet PC, Notebook PC, Desktop PC, Industrial PC and other devices. The terminal device 1 is connected to a file library, and can retrieve files from the file library. The file library can be a storage space on the terminal device 1, a server or a cloud storage space, but is not limited thereto. In this embodiment, the data leakage prevention server 10 further includes an application programming interface 14 (Application Programming Interface), a confidentiality level management unit 16 and an authority management unit 18 . The application program interface 14 can be used for communication and data transmission with the terminal device 1 and the AI server 20 , for example, the data leakage prevention unit 12 on the terminal device 1 can be used to update user rights and set document confidentiality levels. or change data, etc. The confidentiality level management unit 16 is used to manage the confidentiality level labels corresponding to each file. For example, the confidentiality level management unit 16 stores a table of contents such as a table of contents or an index table for recording the confidentiality level labels of each file. The authority management unit 18 is used to manage the authority of each user for browsing, editing or other document processing actions for each confidential level document. Accordingly, the data leakage prevention unit 12 can communicate with the application program interface 14 to download or read the confidentiality level labels corresponding to each file from the confidentiality level management unit 16 , or upload the confidentiality level labels to the confidentiality level management unit 16 , and The authority corresponding to the user of the terminal device 1 is updated from the authority management unit 18 .

In addition, the terminal device 1 can also communicate with the data leakage prevention server 10 through the data leakage prevention unit 12 , and can transmit or upload the file from the terminal device 1 to the data leakage prevention server 10 , and the data leakage prevention server 10 transmits the file to the artificial intelligence analysis module 30 for the artificial intelligence analysis module 30 to analyze the confidentiality level of the document, and the artificial intelligence analysis module 30 The analysis result will be returned to the data leakage prevention server 10, or returned to the data leakage prevention server 10, and then sent back to the data leakage prevention unit 12 on the terminal device 1, and the data leakage The protection server 10 or the data leakage protection unit 12 encrypts or controls the authority of the file according to the analysis result. In this embodiment, the artificial intelligence analysis module 30 includes an application programming interface 32. After the artificial intelligence analysis module 30 completes the analysis of the confidentiality level of the document, the artificial intelligence analysis module 30 can be analyzed through the application programming interface 32. The given confidentiality level label is returned to the confidentiality level management unit 16 of the data leakage prevention server 10 for storage.

The artificial intelligence training module 40 is connected to the artificial intelligence analysis module 30 for training the analysis model used by the artificial intelligence analysis module 30 and for receiving the analysis model of the file through the artificial intelligence analysis module 30 The analysis data generated after the analysis of the confidentiality level, and the analysis model is retrained according to the analysis data.

Please cooperate with FIG. 1 and FIG. 2 together. The method for setting the document confidentiality level of the present invention includes the following steps:

The terminal device 1 provides a file. The manner in which the terminal device 1 provides a file includes but is not limited to: the terminal device 1 downloads a file from the file library, or the user creates a file on the terminal device 1, or the user edits and stores a file on the terminal device 1. . The data leakage protection server 10 or the data leakage protection unit 12 installed on the terminal device 1 will monitor the files stored or executed on the terminal device 1 according to the data leakage protection specification. When the database accesses a file that is subject to the data leakage protection specification, or when the user operates the terminal device 1 to create a file, and wants to store or move the file to the file library, if the file belongs to the data leakage protection For the type monitored by the leakage protection specification, the data leakage protection server 10 will provide the file to the artificial intelligence analysis module 30, and the artificial intelligence analysis module 30 will analyze the confidentiality level of the file according to the analysis model, And provide a first secret level label.

In an embodiment, the data leakage protection unit 12 on the terminal device 1 may transmit the file to the data leakage protection server 10, and then the data leakage protection server 10 sends the file to a manual worker. The intelligent analysis module 30 analyzes and predicts the confidentiality level of the corresponding document. In addition, in another embodiment, the artificial intelligence analysis module 30 further includes a model agent 34. The model agent 34 may be, but is not limited to, an application that is installed on the terminal device 1 for use in Confidentiality level analysis is performed on the file, so that when the terminal device 1 or the data leakage prevention unit 12 cannot connect with the external data leakage protection server 10, the model agent program local to the terminal device 1 can be used. 34. Perform a classification level analysis on the document to provide a corresponding first classification level label. More specifically, the model agent 34 is connected to the artificial intelligence server 20, obtains and updates the analysis model from the artificial intelligence server 20, and can use the analysis model to perform confidentiality level analysis on documents to provide First Class Secret Label.

In addition, the user will also perform a confidentiality level analysis on the file, and provide a second confidentiality level label, and provide the second confidentiality level label to the data leakage prevention server 10, and the data leakage prevention server will 10 Compare whether the first confidentiality level label is the same as the second confidentiality level label. The time point when the user provides the second confidentiality level label may be after or before the artificial intelligence analysis module 30 performs the confidentiality level analysis on the document, for example, in one case, when the user edits a document Or when storing a file, the user can mark the file with a second confidentiality level label, and then the file will be further submitted to the artificial intelligence analysis module 30 for confidentiality level analysis, and a corresponding first confidentiality level label is generated; in In another case, the document is firstly analyzed by the artificial intelligence analysis module 30 for a confidentiality level and a first confidentiality level label is generated, and then submitted to the terminal device 1, so that the user can perform a confidentiality level analysis on the document to provide Second Class Secret Label.

After that, the data leakage prevention server 10 will compare whether the first confidentiality level label is the same as the second confidentiality level label, and if they are the same, mark the file with the first confidentiality level label, and the data leakage prevention server 10 will The server 10 encrypts or controls the file according to the content defined by the first confidentiality level label; 2. The content defined by the confidentiality level label shall be encrypted or subject to authority control. For example, the first confidentiality level label and the second confidentiality level label may be used to classify but not limited to general, secret, secret, extremely secret, absolutely secret and other labels, or the first and second secret level labels The label is used to point to or represent the document confidentiality level set by the user, so that the data leakage prevention server 10 can read and encrypt the document according to the content of the first and second confidentiality level labels or the user can access the document. Permission control of executable and non-executable actions.

In addition, the corresponding analysis data generated after the artificial intelligence analysis module 30 analyzes the document with the confidentiality level will be sent to a training database for storage, and the artificial intelligence training module 40 will be connected to the training database to receive the analysis data. some analysis data, and retrain the analysis model according to the analysis data. In an embodiment, as shown in FIG. 1 , the artificial intelligence training module 40 includes a data collection and labeling tool 42 , a model training tool 44 , a model retraining tool 46 and an error verification tool 48 .

The data collection and labeling tool 42 is used to generate the training data required to build the analytical model. For example, the data collection and labeling tool 42 includes a user interface, and the data collection and labeling tool 42 is connected to a file library , the user can select the files to be analyzed or trained on the interface. The user can select one or more of the files and mark the files with the second confidentiality level label. In addition, the data collection and marking tool 42 also It includes a suggested label interface, which is connected to the terminal device and provides a plurality of baseline models. The user can select one of the baseline models through the suggested label interface and analyze the model by artificial intelligence. The group 30 analyzes the confidentiality level of the document according to the selected benchmark model, and generates at least one suggested confidentiality level label, and then the user decides whether to select the suggested confidentiality level label as the second confidentiality level label, in other words , when the user does not know how to set or determine the second confidentiality level label of the document, the second security level label can be provided with the assistance of the suggested label interface. Wherein, the benchmark model may be a model selected from the document confidentiality level setting system of the same or similar industries, or a model obtained by training based on documents of similar or similar categories. For example, Company A is For the mold manufacturer, the documents that set the confidentiality level are the documents related to the mold design drawings; the company B is also a mold manufacturer or its confidential documents are also related to or similar to the mold design drawings of the company A, and are trained based on the documents of the company A. The benchmark model can be applied to business B, so that business B can use the benchmark model to give a reference to the second confidentiality level label for its documents.

The model training tool 44 is used for receiving the training data to build the analysis model. The algorithms that can be used by the model training tool 44 include but are not limited to: Support Vector Machine (SVM), Multilayer Perceptron (MLP), Gradient Boosting Decision Tree (GBDT), BERT (Bidirectional Encoder Representations from Transformers), LSTM (Long Short-Term Memory) etc. or a combination thereof. Please cooperate with FIG. 3 and FIG. 4 in an embodiment, the model training tool 44 may include but not limited to: automatic word segmentation (word segmentation) module (Word Segmentation Module), word vector module (Word- to-Vector Module), Feature Selection Module, Dimensionality Reduction Module, Weighted Module, and Classification Module. The automatic word segmentation module can perform word segmentation on the words and sentences in the file based on dictionary or thesaurus matching according to different languages (such as Chinese or English, etc.), or perform word segmentation based on word frequency statistics, but not limited thereto; the word vector module The group is used to vectorize the document after word segmentation (word segmentation) into a word vector; the feature selection module may, but is not limited to, use a single letter, two letters or three letters as a feature unit to perform feature selection on the word vector; the described The dimension reduction module can use but not limited to chi-square test (chi-square), one-way independent variance analysis (ANOVA F-value), etc. to reduce the feature dimension; the weighting module can use word frequency-inverse document frequency (TF- IDF) algorithm to evaluate the importance of each word and calculate the weight; the classification module can use but not limited to Support Vector Machine (SVM), Multilayer Perceptron (MLP), Gradient Boosting Decision Tree (GBDT), to build the analytical model. After that, test the analysis model to analyze the security level of the document and generate an effect report. According to the result of the effect report, judge whether the accuracy of the analysis model or other indicators meet the requirements of the user. If the requirements are met, the analysis model is completed. If it does not meet the requirements, continue to adjust the parameters of the model establishment and add training data before training the model. The training method of the analysis model may use, but is not limited to, divide the training data and/or analysis data into a training set and a test set according to a split ratio of 80-20, and then use the modules of the above-mentioned model training tool 44 to assist in the analysis. The model is trained to build the analytical model. In addition, the knowledgeable person can usually choose different analysis model training methods according to the different types of files to be analyzed, and the above description is not limited. The model retraining tool 46 can receive the analysis data stored in the training database to retrain the analysis model. The model retraining tool 46 can adopt a structure similar to that of the model training tool 44, and details are not described herein again.

The error verification tool 48 is used for storing documents with different labels of the first security level and the second security level. In other words, the prediction result of the security level of the document stored by the artificial intelligence analysis module 30 is inconsistent with the security level of the document given by the user. and leave it to the user for subsequent confirmation or improvement of the analysis model. For example, the user can periodically use the error verification tool 48 to check the documents whose first and second confidentiality level labels are inconsistent within a period of time, and confirm whether the first or second confidentiality level label is correct or appropriate. In this case, if the first confidentiality level label is correct or more appropriate, modify the second confidentiality level label to the first confidentiality level label, and store the result in the training database, which can be used for training the analysis model; In the second case, if the second confidentiality level label is correct or more appropriate, the first confidentiality level label is modified to the second confidentiality level label, and the result is input into the model retraining tool 46 with a weight different from that in the first case to Fine-tune the analytical model or retrain the analytical model.

In addition, in one embodiment, the model training tool 44 may also use deep learning technology to create an analysis model, for example, transfer learning (Transfer Learning) may be used to train the analysis model. Please cooperate as shown in Figure 5. The pre-operations for establishing the analysis model include: collecting relevant documents of at least one industry, such as: but not limited to the mold manufacturing industry, machinery industry, electronics industry, medical industry, banking industry, Relevant documents and materials of insurance industry, biotechnology industry, pharmaceutical industry and other industries, the relevant documents and materials may be but not limited to customer information, design drawings, process documents, product formulas, etc. Afterwards, using the language model (Language Model) trained by the general corpus and pre-collected document data of at least one industry, through pre-training (Pre-Training), a pre-training model suitable for the at least one industry is obtained. Afterwards, when the company to be imported into the company is used to set and classify the documents of the company, a suitable pre-training model is selected according to the industry to which the company belongs, or according to the type of documents that the company wants to classify. The company's internal documents and confidential classification and other related information are added. The confidential classification can be derived from the company's confidential document management method, whereby the model training tool 44 can learn the confidentiality of the company's documents. Level features and classification logic of confidentiality level, and fine-tuning the pre-trained model to generate an analysis model suitable for setting the document confidentiality level of the enterprise's documents. In addition, when the document confidentiality level setting system operates for a period of time in the enterprise, the collected files whose first and second confidentiality level labels are inconsistent can be used by the model retraining tool 46 to retrain the analysis model. For example, you can freeze a part of the layers (Layers) in the neural network of the analysis model, such as several low layers, and fine-tune another part of the layers, such as several high layers, in this way to lock the update of specific layer weights, or use Different learning rates are set for each layer of the neural network, and the analysis model is retrained, so as to obtain an analysis model that is more suitable for the enterprise's document confidentiality level setting system.

It is worth mentioning that, in an embodiment, when the artificial intelligence analysis module 30 performs the confidentiality level analysis on the same file again, the first confidentiality level label provided by the artificial intelligence analysis module 30 and the second confidentiality level label provided by the user are provided. When the confidentiality level labels are different, the artificial intelligence training module 40 will compare and extract the difference content between the previous and later versions of the file, and retrain the analysis model according to the difference content. For example, when a user edits or modifies a document whose confidentiality level has been set, and the confidentiality level of the document is changed accordingly, the user will change the second confidentiality level label given to the document, And if the file is judged by the artificial intelligence analysis module 30, the first confidentiality level label given by the artificial intelligence analysis module 30 is different from the second confidentiality level label, indicating the analysis used by the artificial intelligence analysis module 30. The model needs to be updated. Therefore, the artificial intelligence training module 40 will compare and extract the difference content between the previous and later versions of the file, that is, the content of the user's secondary editing and modification of the file, and analyze the model based on the difference content. Retrain to improve the accuracy of using the analytical model to set file confidentiality levels in the future.

Please cooperate with FIG. 6 , in one embodiment, when the data leakage prevention server 10 is monitoring the terminal device 1 , it detects a file that needs to be set with a confidentiality level, for example, it detects that the user operates the terminal device 1 . When a file suspected of containing sensitive data or information is opened, downloaded or stored, or when a file type specified to be tracked or noted in the data leakage protection specification is detected, the data leakage protection server 10 will call The artificial intelligence analysis module 30 performs a confidentiality level analysis on the file, and after the artificial intelligence analysis module 30 provides a corresponding first confidentiality level label, the data leakage prevention server 10 will mark the document with the first confidentiality level label, or establish a data processing link between the file and the first confidentiality level label, and first encrypt or control the file according to the content defined by the first confidentiality level label; in addition, operate the file The user or other users with command and supervision authority will also perform a confidentiality level analysis on the file, and provide a second confidentiality level label, and then the data leakage prevention server 10 will compare the first confidentiality level label with the second confidentiality level label. Similarities and differences of confidentiality level labels, if the first confidentiality level label and the second confidentiality level label are the same, the first confidentiality level label marked in the file is maintained; if the first confidentiality level label and the second confidentiality level label are different, the The file changes are marked as the second confidentiality level label, and then the file is encrypted or authorized according to the content defined by the second confidentiality level label.

Wherein, the confidentiality level tags of the file (the first confidentiality level tag and the second confidentiality level tag) can be stored in the file attributes, and follow the file storage and transfer; in addition, in some embodiments, the confidentiality of the file The level label can also be stored in the data storage space such as the data leakage protection server 10, the terminal device 1 or the data leakage protection unit 12, or stored in a database, and when the confidentiality level label of the corresponding file is updated , the confidentiality level label of the corresponding file in each data storage space can be updated synchronously, and the data leakage prevention unit 12 can go to the above-mentioned data storage space or database to obtain the confidentiality level label of the corresponding file, and according to the content of the confidentiality level label. Encryption or access control.

Please cooperate with FIG. 7A to FIG. 7C. In one embodiment, the confidentiality level label of the file can be stored in the file attribute content of the file, such as the content stored in "document confidentiality level attribute" (an exemplary name). , including the field F1 of the artificial intelligence label and the field F2 of the user label, in the field F1 for filling in the first confidentiality level label given by the artificial intelligence analysis module 30, in the field F2 is used to fill in the label of the second confidentiality level given by the user. FIG. 7A is a schematic diagram of the file confidentiality level attribute of a file that has not been subjected to file confidentiality level analysis. At this time, the file has not been subjected to file confidentiality level analysis. Therefore, its fields F1 and F2 are blank; as shown in FIG. 7B , when the artificial intelligence analysis module 30 gives the first confidentiality level label, for example, the given first confidentiality level label is the code "A", at this time, the field F1 corresponding to the artificial intelligence label will be filled with the corresponding code "A", when the user gives the second confidentiality level label, for example, the given second confidentiality level label is the code "B", at this time, the field F2 of the corresponding user label will be filled with the corresponding code "B", after that, the data leakage prevention server 10 will read the contents of the fields F1 and F2, and make a judgment based on the contents of the fields F1 and F2. In this example, since the first confidentiality level label is the code " A" is not the same as the code "B" of the second confidentiality level label. Therefore, the data leakage prevention server 10 will modify the content of the field F1 of the artificial intelligence label to code "B", and change the field of the user label to the code "B". F2 empties, as well as this time. Afterwards, the security level label code "B" marked on the file is the second security level label given by the user. Therefore, the data leakage prevention server 10 or the data leakage prevention unit on the terminal device 1 12 Encrypts or controls the corresponding file according to the code "B" (second confidentiality level label) retained on the artificial intelligence label.

In addition, please cooperate with FIGS. 8A to 8C , which are basically the same as those shown in FIGS. 7A to 7C , the difference is that the second confidentiality level label given by the user is the same as the first confidentiality level given by the artificial intelligence analysis module 30 . The labels are the same, with the code “A”. Therefore, in FIG. 8C, the AI label field F1 is still the code “A”, which is the first confidentiality level label given by the AI analysis module 30, and the The field F2 of the user label is also cleared after the data leakage prevention server 10 compares the first confidentiality level label with the second confidentiality level label. After that, the document is marked with the confidentiality level label code "A". " is the first confidentiality level label given by the artificial intelligence analysis module 30. Therefore, the data leakage prevention server 10 or the data leakage prevention unit 12 on the terminal device 1 will use the code stored on the artificial intelligence label according to the code. "A" (first secret level label) encrypts or controls the corresponding file. In addition, it should be mentioned that the aforementioned codes "A" and "B" are only for illustrative purposes, and are not the contents of the actual confidentiality level label.

The advantage of clearing the field F2 corresponding to the user label that stores the second confidentiality level label is that when the document marked with the confidentiality level label is edited by the user again and the content is changed, the second confidentiality level label is stored. The field F2 of the user label of the confidentiality level label is blank, therefore, the data leakage prevention server 10 will require the user to confirm the confidentiality level of the document again, so as to give a new second confidentiality level label, and then communicate with The first confidentiality level label given by the artificial intelligence analysis module 30 is compared, so as to avoid evaluating the confidentiality level of the document with the old second confidentiality level label after the document has been edited and the content has changed. When this happens, the document confidentiality level can be reassessed in real-time and on-the-fly for redacted documents. In other words, the aforementioned action of clearing the field storing the second confidentiality level label after comparing the first confidentiality level label with the second confidentiality level label will help to perform the document security classification of the re-edited document. For evaluation, the data leakage prevention server 10 can check the status of the field of the user label to determine whether to submit a document to the user for a confidential level analysis, so as to allow the user to provide a real-time second confidential level label.

The system and method for setting the document confidentiality level provided by the present invention can not only classify the document confidentiality level according to the pre-learned artificial intelligence analysis module, but also return the document confidentiality level set by the user after the document is generated or edited. And when there is a discrepancy between the document confidentiality level marked by the user and the document confidentiality level label marked by the artificial intelligence analysis module, the document will be included in the data relearned by the artificial intelligence analysis module, and the document can be compared with the data based on the document. The differences between the previous versions of the document retrain the AI analysis module, thereby improving the accuracy of automatic classification of future AI analysis modules. In addition, when the document is modified or re-edited, the document can also be re-examined to see if there is a need to change the document's confidentiality level, and the document's confidentiality level can be set immediately and accurately.

In order to increase the accuracy of document confidentiality setting by the system and method for setting document confidentiality levels, please cooperate with FIG. 9 to provide a document confidentiality level management system 100 according to an embodiment of the present invention, which includes a plurality of the above-mentioned document confidentiality level setting systems 110 and a Federated Learning System 120.

The document confidentiality level setting systems 110 are respectively used in a plurality of different units of the same industry category, such as different companies in the same industry. The federated learning system 120 is connected to each of the document confidentiality level setting systems 110 . In this embodiment, the federated learning system 120 includes a federated learning server 122 . Smart server 20 is connected. The AI server 20 further includes a model parameter encryption module 50 and a communication module 52 . The federated learning server 122 includes a data quality assessment module 122a, a data filtering module 122b, a feature aggregation module 122c, a model parameter update storage unit 122d, a model storage unit 122e and a model delivery module 122f.

Please cooperate with FIG. 9 and FIG. 10 to describe the document confidentiality level management method of the present invention, which includes the steps performed by each of the document confidentiality level setting systems 110 and the steps performed by the federated learning system 120, wherein:

The steps performed by the document confidentiality level setting system 110 are based on the document confidentiality level setting method shown in FIG. 2 , and further include encrypting a model parameter of the analysis model into an encrypted parameter and sending it to the federated learning system 120 , For example, taking a neural network-like road algorithm as an example, the model parameters may include multiple parameters, including gradients, weights, deviations, the number of layers of the neural network, loss function, size of convolution kernel, learning rate, momentum, and number of iterations etc., in practice, the model parameters may include at least one parameter. In this embodiment, the model parameters of the analysis model are encrypted by the model parameter encryption module 50 using an encryption algorithm, for example, advanced encryption standard (Advanced Encryption Standard, AES), RSA encryption algorithm and other algorithms can be used, but Not limited to this, the communication module 52 of the AI server 20 then transmits the encrypted encryption parameters to the federated learning server 124 .

The federated learning system 120 performs the following steps:

Receive the encryption parameters transmitted from the document confidentiality level setting system 110 . In this embodiment, the federated learning server 124 receives the encryption parameters.

The encryption parameters are aggregated and an update parameter is generated.

In this embodiment, the federated learning server 124 processes the received encryption parameters under homomorphic encryption, including screening and aggregation. The data quality evaluation module 122a and the data filtering module 122b are used for screening encryption parameters, and the data quality evaluation module 122a performs quality evaluation processing on the received encryption parameters to determine whether each encryption parameter can be used for aggregation, If one of the encryption parameters (eg gradient) deviates from the confidence interval, it is not used for aggregation to avoid affecting the update parameter. The data filtering module 122b filters the received encryption parameters to find features in the encryption parameters that can be used for aggregation. The feature aggregation module 122c aggregates the encrypted parameters after the aforementioned processing into the update parameters, which are stored in the model storage unit 122e. In one embodiment, the encryption parameters can be further obfuscated.

The update parameters are transmitted to each of the document security level setting systems 110 . In this embodiment, the model dispatching module 122f transmits the updated parameters in the model storage unit 122e to the communication module 52 of the artificial intelligence server 20 of each document confidentiality level setting system 110 .

Each of the document security level setting systems 110 further performs the following steps:

Each of the document security level setting systems 110 receives the update parameter.

Each of the artificial intelligence analysis modules 30 will update the analysis model according to the update parameter. In this embodiment, each of the artificial intelligence analysis modules 30 decrypts the update parameters, and performs operations based on the decrypted update parameters and the original model parameters of the analysis model, such as summing or averaging the same parameters, to update the analytical model. In this embodiment, after receiving the update parameter, the communication module 52 determines whether to transmit the update parameter to the artificial intelligence analysis module 30 to update the analysis model. If so, the communication module 52 drives the artificial intelligence analysis module 30 to update the analysis model. If the user thinks that the original analysis model is accurate enough, it can be set in the communication module 52 not to transmit the updated parameters to the artificial intelligence analysis module 30 .

Afterwards, when each of the document confidentiality level setting systems 110 executes the document confidentiality level setting method again, it can analyze the document confidentiality level according to the updated analysis model.

With the above-mentioned document confidentiality level management system 100 and method, the amount of data for model training can be increased, so as to make up for the shortcoming of insufficient amount of data for model training in each of the document confidentiality level setting systems in each unit, thereby improving artificial intelligence analysis The accuracy of the judgment of the module 30 . It is worth mentioning that each of the document confidentiality level setting systems 110 encrypts the model parameters and uploads them to the federated learning system 120 instead of uploading the analysis model. Therefore, even after obtaining the encrypted parameters, it is difficult to crack, even if the encrypted parameters are encrypted. Even if it is deciphered, the analysis model cannot be obtained. In this way, the leakage of confidential documents in each unit can be effectively avoided, and the data security of documents in the document confidentiality level setting system 110 can be improved.

In addition, when using the updated analysis model to perform the document confidentiality level setting method, the artificial intelligence training module 40 also receives the analysis data generated by the artificial intelligence analysis module 30 after the document has been analyzed for the confidentiality level, and based on the The analysis data retrains the analysis model. In other words, after the updated analysis model is retrained by the document confidentiality level setting system 110 of each unit, an analysis model more suitable for each unit can be generated, which can further improve the accuracy of the judgment of the artificial intelligence analysis module 30 . Afterwards, each of the document confidentiality level setting systems 110 may further encrypt the model parameters of the retrained analysis model to form the encrypted parameters and then transmit them to the federated learning system for further aggregation to generate an updated model.

Please cooperate with FIG. 11 and FIG. 12 to provide a document security level management system 101 according to another embodiment of the present invention. In this embodiment, the document security level setting systems 110 are divided into a plurality of categories C, and the categories C can be, for example, Different industry categories, such as manufacturing, medical, financial. Each of the categories includes a plurality of the document security level setting systems 110 , at least two document security level setting systems 110 . Each of the document security level setting systems 110 is used in a plurality of different units for each category C.

In this embodiment, the federated learning system 130 includes a plurality of federated platforms 132 and a federated learning server 134, each of the federated platforms 132 is connected to the file confidentiality level setting system 110 of each type, and the federated learning server 134 is connected to the federated learning server 134. These federated platforms 132 are connected. Each of the federation platforms 132 includes the data quality assessment module 122a, the data filtering module 122b, the feature aggregation module 122c, and the model parameter update storage unit 122d. The federated learning server 134 includes a model optimization update module 134a, the model storage unit 122e, and the model dispatch module 122f.

The document confidentiality level management method of this embodiment is shown in FIG. 13 , which has roughly the same steps as those in FIG. 10 , with the following differences:

The encryption parameters of the document confidentiality level setting system 110 for each of the Class Cs are received by each of the federated platforms 132 .

Each of the federated platforms 132 aggregates the received encryption parameters of the document confidentiality level setting system 110 of each category into a relay parameter and then transmits it to the federated learning server 134 . In this embodiment, the processing performed by each module of the federation platform 132 is the same as that of the previous embodiment, the difference is that the feature aggregation module 122c aggregates the processed encryption parameters into the relay parameters, and Stored in the model storage unit 122e.

The federated learning server 134 optimizes and updates the relay parameters under homomorphic encryption to generate the updated parameters. Since the confidential words or sentences in the documents of each category C are slightly different, in this embodiment, the model optimization and update module 134a optimizes and updates the relay parameters transmitted from the federated platforms. The optimization and update may For example, some parameters among the relay parameters of the similar category C are selected and re-aggregated to generate update parameters suitable for these categories and store them in the model storage unit 122e. Different categories C respectively generate a plurality of different update parameters suitable for each category C. The model dispatching module 122f transmits the updated parameters in the model storage unit 122e to the communication module 52 of the AI server 20 of each of the document confidentiality level setting systems 110 . If there are a plurality of update parameters, the model dispatching module 122f transmits the update parameters to the document security level setting system 110 of the corresponding category C, respectively.

The encryption parameters transmitted from the file confidentiality level setting system 110 of each category C by each of the federation platforms 132 may include performing quality assessment processing, obfuscation processing, and aggregation into the relay parameters under homomorphic encryption.

This embodiment can be applied to aggregate the encryption parameters of the document confidentiality level setting system 110 of different industry categories to generate update parameters, which can not only increase the amount of training data of each unit, but also utilize the confidential word of the relevant industry category. words or phrases as judged confidential documents. Similarly, leakage of confidential documents in each unit can also be effectively avoided, so as to improve the data security of documents in each document confidentiality level setting system 110 .

Please cooperate with FIG. 14 to provide a file confidentiality level management system 102 according to another embodiment of the present invention. In this embodiment, it is based on the aforementioned file confidentiality level management system 100 . The difference is that the federated learning system 140 includes a plurality of Federated learning servers 142, each of which is connected to the document confidentiality level setting system 110 of each category C, and each category C includes at least two document security level setting systems 110. The document confidentiality level management method of this embodiment is shown in FIG. 15 , which has roughly the same steps as those in FIG. 10 , with the following differences:

In the federated learning system 140, the federated learning servers 142 receive the encryption parameters of the document confidentiality level setting system 110 of the category C; The encrypted parameters are aggregated into a relay parameter, and two of the federated learning servers 142 further transmit their relay parameters to each other and combine the relay parameters from the other party with the document confidentiality level from each category. The encryption parameters of the setting system 110 are aggregated into an update parameter; then, each of the federated learning servers 142 transmits each of the updated parameters to each of the document confidentiality level setting systems 110 of the corresponding category C.

In the document confidentiality level management system 102 of this embodiment, it can be applied to aggregate encryption parameters and relay parameters of similar categories to generate update parameters, which can not only increase the amount of training data of each unit, but also can utilize related industry categories Confidential words or phrases used as judged confidential documents. Similarly, leakage of confidential documents in each unit can also be effectively avoided, so as to improve the data security of documents in each document confidentiality level setting system 110 .

According to the above, the document confidentiality level management system and method of the present invention can effectively retrain the analysis models of different document confidentiality level setting systems 110 into a more accurate analysis model, which overcomes the problem in the single document confidentiality level setting system 110 . In order to improve the accuracy of judgment, the amount of training data is insufficient.

Since the data transmitted by the document confidentiality level setting system 110 to the federated learning systems 120, 130, and 140 are the model parameters of the analysis model and the model parameters are transmitted after being encrypted, instead of directly transmitting the file or the analysis model, it can be effectively avoided. In the case of leakage of confidential documents, the data security of documents in the document confidentiality level setting system 110 is improved.

In addition, the document confidentiality level management systems 100, 101, 102 can also be applied in the same enterprise, and the document confidentiality level setting systems 110 are respectively in a plurality of different departments of the same enterprise, for example, including the research and development department, the business department, the accounting department Department, Marketing Department, Design Department and other manufacturing departments. Users in the same department have similar standards for identifying confidentiality levels, and users in different departments have different standards for identifying confidentiality levels. Therefore, each department uses its own document confidentiality level setting system 110 for training, so as to avoid users in different departments from using a single document. A document confidentiality level setting system 110 and the difference in the identification of document confidentiality levels are too large, resulting in an inaccurate analysis model of a single document confidentiality level setting system 110 , for example, the weight of parameters is reduced, and the feature disorder is increased. Through the federated learning system 120, the parameters of each analysis model can be aggregated, or selected part of the parameters can be aggregated, and then sent back to the document confidentiality level setting system 110 of each department, so as to improve the accuracy of each analysis model.

The above is only a feasible embodiment of the present invention, wherein, in the document confidentiality level setting system, marking the file with the first confidentiality level label or the second confidentiality level label may be, but not limited to, marking the confidentiality level label with the corresponding document. In the field of file attributes, or marked or recorded in a table of contents or index table, or stored in the data storage space such as the data leakage protection server, terminal device or data leakage protection unit, or Stored in a database for access by the data leakage prevention server. When the data leakage prevention server needs to encrypt or control the document, it is derived from the aforementioned file attribute fields, forms, and data. Read the confidentiality level label corresponding to the file in the storage space or database. All equivalent changes made by applying the description of the present invention and the scope of the patent application should be included in the patent scope of the present invention.

1: Terminal device 10: Data leakage protection server 12: Data leakage prevention unit 14: API 16: Confidentiality level management unit 18: Rights Management Unit 20: AI Server 30: Artificial Intelligence Analysis Module 32: API 34: Model Agent 40: Artificial Intelligence Training Module 42: Data collection and tagging tools 44: Model training tools 46: Model Retraining Tool 48: Error Validation Tool 100: Document Confidentiality Level Management System 110: Document Confidentiality Level Setting System 50: Model parameter encryption module 52: Communication module 120: Federated Learning Systems 124: Federated Learning Server 122a: Data Quality Assessment Module 122b: Data filtering module 122c: Feature Aggregation Module 122d: Model parameter update storage unit 122e: Model storage unit 122f: Model Delivery Module 101: Document Confidentiality Level Management System 130: Federated Learning Systems 132: Federation Platform 134: Federated Learning Server 134a: Model optimization update module 102: Document Confidentiality Level Management System 140 Federated Learning Systems 142 Federated Learning Server C: Category

1 is a block diagram of a system for setting document confidentiality levels according to an embodiment of the present invention; 2 is a flowchart of a method for setting a document confidentiality level according to an embodiment of the present invention; 3 is an architectural diagram of a model training tool according to an embodiment of the present invention; FIG. 4 is a training flow chart of an analysis model according to an embodiment of the present invention; Fig. 5 is the training flow chart of the analysis model of another embodiment of the present invention; 6 is a flowchart of a method for setting a document confidentiality level according to another embodiment of the present invention; 7A to 7C are schematic diagrams of document confidentiality level label marking according to an embodiment of the present invention; 8A to 8C are schematic diagrams of document confidentiality level label marking according to an embodiment of the present invention; 9 is a block diagram of a document confidentiality level management system according to an embodiment of the present invention; 10 is a flowchart of a method for managing a document confidentiality level according to an embodiment of the present invention; 11 is a block diagram of a document confidentiality level management system according to another embodiment of the present invention; Fig. 12 is a block diagram of the federated learning system in Fig. 11; 13 is a flowchart of the method for setting a document confidentiality level according to the above-mentioned embodiment of the present invention; 14 is a block diagram of a document confidentiality level management system according to another embodiment of the present invention; FIG. 15 is a flowchart of the method for setting a document confidentiality level according to the above-mentioned embodiment of the present invention.

100: Document Confidentiality Level Management System

110: Document Confidentiality Level Setting System

20: AI Server

30: Artificial Intelligence Analysis Module

40:Artificial intelligence training module"

50: Model parameter encryption module

52: Communication module

120: Federated Learning Systems

122: Federated Learning Server

122a: Data Quality Assessment Module

122b: Data filtering module

122c: Feature Aggregation Module

122d: Model parameter update storage unit

122e: Model storage unit

122f: Model Delivery Module

Claims (15)

A document confidentiality level management system, comprising: a plurality of document confidentiality level setting systems, each of the document confidentiality level setting systems including an artificial intelligence analysis module and a data leakage protection server, wherein: the artificial intelligence analysis module It is used for receiving a file, analyzing the confidentiality level of the file according to an analysis model, and providing a first confidentiality level label according to the analysis result; the data leakage prevention server receives the first confidentiality level label and associates it with a The terminal device is connected, the terminal device is used for user operation and provides a second confidentiality level label, the data leakage prevention server compares whether the first confidentiality level label and the second confidentiality level label are the same, if they are the same , the data leakage prevention server marks the file with the first confidentiality level label, and encrypts or controls the file according to the content defined by the first confidentiality level label; if not, the data leakage The protection server marks the file with the second confidentiality level label, and encrypts the file or controls the authority according to the content defined by the second confidentiality level label; a federated learning system, and each file confidentiality level setting The system is connected, and each of the document confidentiality level setting systems encrypts a model parameter of each of the analysis models to form an encrypted parameter and transmits it to the federated learning system; the federated learning system aggregates these encrypted parameters and generates an update parameter, and the The update parameter is sent to each of the document confidentiality level setting systems, and each of the artificial intelligence analysis modules updates the analysis model according to the update parameter; wherein, each of the document confidentiality level setting systems includes an artificial intelligence training module, the artificial intelligence training module The module is connected with the artificial intelligence analysis module, and is used for receiving analysis data generated by the artificial intelligence analysis module after the confidential level analysis of the document, and retraining the analysis model according to the analysis data; Wherein, the artificial intelligence training module includes a data collection and labeling tool, an error verification tool, a model training tool and a heavy training tool, the data collection and labeling tool is used to generate training data required for establishing the analysis model; the The error verification tool is used for storing the file with the first confidentiality level label and the second confidentiality level label different; the model training tool is used for establishing the analysis model according to the training data; the retraining tool is used for the analysis according to the analysis data The model is retrained. The document confidentiality level management system as claimed in claim 1, wherein the document confidentiality level setting systems are divided into a plurality of categories, each of which includes at least two document confidentiality level setting systems; the federated learning system includes a plurality of federated The platform is connected with a federated learning server, each of the federated platforms is connected with the document confidentiality level setting system of each type, the federated learning server is connected with the federated platforms; each of the federated platforms receives the document security level setting system of each type The encrypted parameters are aggregated into a relay parameter and sent to the federated learning server; the federated learning server optimizes and updates the relay parameters to generate at least one update parameter. The document confidentiality level management system according to claim 2, wherein the federated learning server generates a plurality of the update parameters, each of the update parameters corresponds to each of the categories, and each of the update parameters is transmitted to the corresponding of the categories respectively The document confidentiality level setting system. The document confidentiality level management system according to claim 1, wherein the document confidentiality level setting systems are divided into a plurality of categories, and each of the categories includes at least two of the document confidentiality level setting systems; wherein the federated learning system includes a plurality of A federated learning server, each federated learning server receives the encryption parameters of the file confidentiality level setting system of each category and aggregates them into a relay parameter, and two of the federated learning servers further convert their respective relay parameters to each other and relay parameters from each other with files from each category The encryption parameters of the security level setting system are aggregated into the updated parameters and then sent to the file security level setting systems of the corresponding category. The document confidentiality level management system of claim 1, wherein each of the document confidentiality level setting systems comprises encrypting the retrained model parameters of the analysis model to form the encrypted parameters and then transmit them to the federated learning system. The document confidentiality level management system of claim 1, wherein the data collection and labeling tool includes a suggested labeling interface, the suggested labeling interface is connected to the terminal device, and provides a plurality of reference models, through which the user can The terminal device selects one of the benchmark models, and the artificial intelligence analysis module performs a confidentiality level analysis on the document according to the benchmark model, so as to provide the second confidentiality level label. A document confidentiality level management system, comprising: a plurality of document confidentiality level setting systems, each of the document confidentiality level setting systems including an artificial intelligence analysis module and a data leakage protection server, wherein: the artificial intelligence analysis module It is used for receiving a file, analyzing the confidentiality level of the file according to an analysis model, and providing a first confidentiality level label according to the analysis result; the data leakage prevention server receives the first confidentiality level label and associates it with a The terminal device is connected, the terminal device is used for user operation and provides a second confidentiality level label, the data leakage prevention server compares whether the first confidentiality level label and the second confidentiality level label are the same, if they are the same , the data leakage prevention server marks the file with the first confidentiality level label, and encrypts or controls the file according to the content defined by the first confidentiality level label; if not, the data leakage The protection server marks the document with the second confidentiality level label, and performs encryption or authority control on the document according to the content defined by the second confidentiality level label; a federated learning system, connected with each of the document confidentiality level setting systems, each of the document security level setting systems encrypts a model parameter of each of the analysis models to form an encrypted parameter and transmits it to the federated learning system; the federated learning system sets the These encryption parameters are aggregated to generate an update parameter, and the update parameter is sent to each of the document confidentiality level setting systems, and each of the artificial intelligence analysis modules updates the analysis model according to the update parameter; wherein, the data leakage prevention server Including a data leakage protection unit, the data leakage protection unit is installed on the terminal device, and communicates with the data leakage protection server, the data leakage protection unit is used for uploading the file on the terminal device to the the data leakage prevention server, and the data leakage prevention server transmits the document to the artificial intelligence analysis module to analyze the document with a confidentiality level, and the data leakage prevention unit is based on the document The content defined by the marked first security level label or the second security level label is encrypted or authorized to control the file. A document confidentiality level management system, comprising: a plurality of document confidentiality level setting systems, each of the document confidentiality level setting systems including an artificial intelligence analysis module and a data leakage protection server, wherein: the artificial intelligence analysis module It is used for receiving a file, analyzing the confidentiality level of the file according to an analysis model, and providing a first confidentiality level label according to the analysis result; the data leakage prevention server receives the first confidentiality level label and associates it with a The terminal device is connected, the terminal device is used for user operation and provides a second confidentiality level label, the data leakage prevention server compares whether the first confidentiality level label and the second confidentiality level label are the same, if they are the same , the data leakage prevention server marks the file with the first confidentiality level label, and encrypts or controls the file according to the content defined by the first confidentiality level label; if not, the data leakage Guard Servo The device marks the file with the second confidentiality level label, and encrypts or controls the file according to the content defined by the second confidentiality level label; a federated learning system is connected to each of the document confidentiality level setting systems , each of the document confidentiality level setting systems encrypts a model parameter of each of the analysis models to form an encrypted parameter and transmits it to the federated learning system; the federated learning system aggregates these encrypted parameters and generates an update parameter, and the update The parameters are sent to each of the document confidentiality level setting systems, and each of the artificial intelligence analysis modules updates the analysis model according to the update parameters; wherein, the artificial intelligence analysis module includes a model agent program, and the model agent program is installed on the terminal On the device, it is used to analyze the confidentiality level of the files on the terminal device according to the analysis model. A file confidentiality level management method is applied to a file confidentiality level management system, the file confidentiality level management system includes a plurality of file confidentiality level setting systems and a federated learning system, each of the file confidentiality level settings includes a data leakage protection server the device and an artificial intelligence analysis module, each of the document confidentiality level setting systems is connected with the federated learning system, and the method includes the following steps: each of the document confidentiality level setting systems performs the following steps: A1. A terminal device provides a document; A2. The document is analyzed by the artificial intelligence analysis module according to an analysis model for the confidentiality level of the document, and the artificial intelligence analysis module provides a first confidentiality level label to the data leakage prevention server according to the analysis result; A3. The document is analyzed by a user for the confidentiality level, and a second confidentiality level label is provided to the data leakage prevention server; A4. The data leakage prevention server compares the first confidentiality level label with the The similarities and differences of the second classification label; when the first classification label and the second classification label When signing the same, the data leakage prevention server marks the document with the first-class security label, and encrypts or controls the file according to the content defined by the first-class security label; when the first-class security label is When the class label is different from the second class label, the data leakage prevention server marks the document with the second class label, and the data leakage protection server defines the content according to the second class label, Encrypt or control the authority of the file; A5. Encrypt a model parameter of the analysis model to form an encrypted parameter and transmit it to the federated learning system; the federated learning system performs the following steps: B1. Receive the document confidentiality level settings Encryption parameters transmitted from the system; B2, aggregate these encryption parameters and generate an update parameter; B3, transmit the update parameter to each of the document security level setting systems; each of the document security level setting systems further performs the following steps: A6, each of the document confidentiality level setting systems receives the update parameter; A7, each of the artificial intelligence analysis modules updates the analysis model according to the update parameter; wherein, each of the document confidentiality level setting systems includes an artificial intelligence training module, the The method further includes: the artificial intelligence training module receives analysis data generated by the artificial intelligence analysis module performing a confidentiality level analysis on the document, and retrains the analysis model according to the analysis data; further comprising the following steps : When the artificial intelligence analysis module performs a confidential level analysis on the same file again, and the provided first confidentiality level label is different from the second confidential level label provided by the user, the artificial intelligence training module will compare The difference content between the previous and later versions of the file is extracted, and the analysis model is retrained according to the difference content. The document confidentiality level management method as claimed in claim 9, wherein the document confidentiality level setting systems are divided into a plurality of categories, each of which includes at least two document confidentiality level setting systems; the federated learning system includes a plurality of federated The platform is connected to a federated learning server, each of the federated platforms is connected to the file confidentiality level setting system of each category, and the federated learning server is connected to the federated platforms; wherein: in step B1, each of the federated platforms receives the The encryption parameters of the file confidentiality level setting system of the category; in step B2, each federated platform aggregates the received encryption parameters of the file confidentiality level setting system for each category into a relay parameter and transmits it to the federated learning server ; The federated learning server optimizes and updates these relay parameters to generate at least one of the updated parameters. The document confidentiality level management method as described in claim 10, wherein in step B2, a plurality of the update parameters are generated, and each of the update parameters corresponds to each of the categories; and in step B3, each of the update parameters is transmitted to the corresponding The document security level setting system for each of the categories. The document confidentiality level management method according to claim 9, wherein the document confidentiality level setting systems are divided into a plurality of categories, and each of the categories includes at least two document confidentiality level setting systems; wherein the federated learning system includes a plurality of A federated learning server; wherein in step B1, each federated learning server receives the encryption parameters of the document confidentiality level setting system of each type; wherein in step B2, each federated learning server receives each type of document confidentiality level The encryption parameters of the system are set and aggregated into a relay parameter, and two of the federated learning servers further transmit their relay parameters to each other and combine the relay parameters from each other with files from each category Confidentiality level setting system encryption parameters Aggregate into the update parameter; wherein in step B3, each federated learning server transmits each update parameter to each of the document confidentiality level setting systems of the corresponding category. The document confidentiality level management method as claimed in claim 9 includes that each document confidentiality level setting system encrypts the retrained model parameters of the analysis model to form the encrypted parameters and transmits them to the federated learning system. A file confidentiality level management method is applied to a file confidentiality level management system, the file confidentiality level management system includes a plurality of file confidentiality level setting systems and a federated learning system, each of the file confidentiality level settings includes a data leakage protection server the device and an artificial intelligence analysis module, each of the document confidentiality level setting systems is connected with the federated learning system, and the method includes the following steps: each of the document confidentiality level setting systems performs the following steps: A1. A terminal device provides a document; A2. The document is analyzed by the artificial intelligence analysis module according to an analysis model for the confidentiality level of the document, and the artificial intelligence analysis module provides a first confidentiality level label to the data leakage prevention server according to the analysis result; A3. The document is analyzed by a user for the confidentiality level, and a second confidentiality level label is provided to the data leakage prevention server; A4. The data leakage prevention server compares the first confidentiality level label with the Similarities and differences of the second confidentiality level label; when the first confidentiality level label is the same as the second confidentiality level label, the data leakage prevention server marks the document with the first confidentiality level label, and according to the first confidentiality level The content defined by the tag is encrypted or authorized to control the file; when the first confidentiality level tag is different from the second confidentiality level tag, the data leakage prevention server marks the document with the second confidentiality level tag , by the The material leakage prevention server encrypts or controls the authority of the file according to the content defined by the second confidentiality level label; A5. Encrypt a model parameter of the analysis model to form an encrypted parameter and transmit it to the federated learning system ; The federated learning system executes the following steps: B1, receive the encryption parameters transmitted from the document confidentiality level setting system; B2, aggregate these encryption parameters and generate an update parameter; B3, transmit the update parameter to each of the A document security level setting system; each of the document security level setting systems further executes the following steps: A6, each of the document security level setting systems receives the update parameter; A7, each of the artificial intelligence analysis modules updates the analysis model according to the update parameter; Wherein, after step A2, the data leakage prevention server marks the file with the first confidentiality level label, and encrypts or controls the authority of the file according to the content defined by the first confidentiality level label; in In step A4, when the second confidentiality level label is different from the first confidentiality level label, the data leakage prevention server changes the marking of the file to the second confidentiality level label, and according to the second confidentiality level label The defined content is encrypted or permission control is performed on the file. A file confidentiality level management method is applied to a file confidentiality level management system, the file confidentiality level management system includes a plurality of file confidentiality level setting systems and a federated learning system, each of the file confidentiality level settings includes a data leakage protection server the device and an artificial intelligence analysis module, each of the document confidentiality level setting systems is connected with the federated learning system, and the method includes the following steps: each of the document confidentiality level setting systems performs the following steps: A1. A terminal device provides a document; A2. The document is analyzed by the artificial intelligence analysis module according to an analysis model for the confidentiality level of the document, and the artificial intelligence analysis module provides a first confidentiality level label to the data leakage prevention server according to the analysis result; A3. The document is analyzed by a user for the confidentiality level, and a second confidentiality level label is provided to the data leakage prevention server; A4. The data leakage prevention server compares the first confidentiality level label with the Similarities and differences of the second confidentiality level label; when the first confidentiality level label is the same as the second confidentiality level label, the data leakage prevention server marks the document with the first confidentiality level label, and according to the first confidentiality level The content defined by the tag is encrypted or authorized to control the file; when the first confidentiality level tag is different from the second confidentiality level tag, the data leakage prevention server marks the document with the second confidentiality level tag , the data leakage prevention server encrypts or controls the file according to the content defined by the second confidentiality level label; A5, encrypts a model parameter of the analysis model to form an encryption parameter and transmits it to the A federated learning system; the federated learning system executes the following steps: B1, receiving encryption parameters from the document confidentiality level setting system; B2, aggregating these encryption parameters and generating an update parameter; B3, transmitting the update parameter to each of the document security level setting systems; each of the document security level setting systems further performs the following steps: A6. Each of the document security level setting systems receives the update parameter; A7, each of the artificial intelligence analysis modules updates the update parameter according to the update parameter. Analyzing model; wherein, in step A3, the user selects one of the benchmark models among several benchmark models, and the artificial intelligence analysis module selects the benchmark model according to the user's selection The quasi-model analyzes the confidentiality level of the document, and the user provides the second confidentiality level label according to the analysis result.

Images