Intelligent file classification and marking management system

Abstract

The present application relates to the technical field of data security, in particular to a kind of file intelligent fixed encryption marking tube control system.It includes heterogeneous information network construction module, for personnel, document, project, equipment entity and its associated relationship are constructed as dynamic update heterogeneous information network;Implication point mining module, for mining the implication point of secret entity in electronic file with implication association, and verifying its key contribution degree, generating fixed encryption suggestion;Adversarial evolution watermark embedding module, for encoding fixed encryption result into watermark information that is symbiotic with file content by generative adversarial watermark network;Stress enhancement protection module, for constructing user behavior graph, detecting behavior deviation and triggering dynamic watermark enhancement;Strobe coding traceability module, for embedding strobe coding to realize leakage traceability when the highest level protection is embedded.The present application realizes the complete closed loop of intelligent fixed encryption, anti-attack marking, dynamic protection and leakage traceability, effectively improves the security of secret-related electronic files.

Description

Technical Field

[0001] This invention relates to the field of data security technology, and more specifically, to a document intelligent classification and classification control system. Background Technology

[0002] With the deepening of information technology development, the number of electronic documents generated by various organizations has exploded, including a large amount of classified information and sensitive data. How to accurately and efficiently classify and label these electronic documents, and implement effective security management throughout their entire lifecycle, has become a critical issue that urgently needs to be addressed in the field of information security.

[0003] Traditional electronic document confidentiality control relies primarily on manual confidentiality assessment and static access control. In recent years, some systems have begun to incorporate natural language processing (NLP) technology to assist in confidentiality assessment; however, this often involves shallow keyword matching, failing to understand contextual semantics and implicit sensitive relationships. While digital watermarking technology has been attempted for confidentiality assessment, traditional watermarking suffers from a contradiction between robustness and imperceptibility, making it vulnerable to common attacks such as screenshots, print scans, and other malicious activities. Behavioral auditing is mostly based on post-event log analysis, lacking real-time risk awareness and proactive intervention capabilities. Therefore, this paper proposes an intelligent document confidentiality control system.

[0004] The existing technology has the following technical defects, specifically:

[0005] 1. The classification relies on shallow matching and cannot discover hidden risks: Existing technologies mostly use keyword matching or shallow semantic analysis, which can only identify sensitive words directly mentioned in the document. They cannot uncover the multi-hop hidden relationships between the document content and the classified entities (personnel, projects, equipment), resulting in a large number of documents that are "not directly mentioned but are actually classified" being missed. Moreover, the classification process lacks interpretability and cannot provide specific evidence to support the classification conclusion.

[0006] 2. Loose coupling between security classification and content; watermarks are easy to remove but difficult to trace: Traditional security classification methods use header markings or static watermarks, separating the security classification from the document content. Once the document is screenshotted, formatted, or printed and scanned, the security classification information is lost, leading to complete loss of control after secondary dissemination. Existing watermarking technologies have an inherent contradiction between robustness and imperceptibility, and are all passive defenses, unable to cope with new attack methods.

[0007] 3. Static and crude protection, lacking real-time risk perception and intervention capabilities: The existing system adopts a static permission model of "one-time authorization, long-term validity", which cannot perceive the dynamic changes in user operation behavior. It lacks the ability to identify and actively intervene in high-risk behaviors such as malicious operation by legitimate users, identity impersonation, and screen capture in real time, and all of them are audited after the fact, which cannot prevent leakage during the risk process. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent document classification and control system to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention aims to provide a document intelligent classification and classification control system, comprising: a heterogeneous information network construction module, used to construct a dynamically updated heterogeneous information network from personnel entities, document entities, project entities, equipment entities and their relationships within the target organization; the heterogeneous information network includes multiple types of nodes and multiple types of edges, and uses a time-series decay factor to dynamically adjust the weights of the edges.

[0010] The hidden key information mining module is used to perform multi-hop association reasoning on the heterogeneous information network, to mine key information in the electronic document content that is not directly mentioned but has an implicit association with known classified entities, and to verify the key contribution of the key information to the document's classification level, and to generate classification suggestions and classification basis.

[0011] An adversarial evolution watermark embedding module is used to construct a generative adversarial watermarking network. The generative adversarial watermarking network includes a watermark generation network and an attack simulation network. Through continuous adversarial training of the two, watermark information coexists with the feature space of the electronic document content, and the encryption result is encoded and embedded into the watermark information.

[0012] The stress-enhanced protection module is used to collect user operation behavior data in real time and construct a user behavior map. When the deviation between the current operation behavior and the user's historical behavior map exceeds a preset threshold, the dynamic watermark enhancement mode of the adversarial evolution watermark embedding module is triggered to improve the embedding strength and robustness of the watermark.

[0013] As a further improvement to this technical solution, the heterogeneous information network construction module includes: a multi-source entity access unit, used to access data from the personnel information system, document management system, project management system, and equipment monitoring system within the organization, and to extract personnel nodes, document nodes, project nodes, and equipment nodes.

[0014] The relationship extraction unit is used to extract various types of relationships between entities, including the writing relationship between documents and people, the citation relationship between documents, the participation relationship between people and projects, the printing / sending relationship between documents and devices, and the chronological relationship between events and time.

[0015] The dynamic weight update unit is used to assign an initial weight to each associated edge and periodically decay the weight of the edge according to the time decay function. At the same time, when a new business interaction is detected, the weight of the corresponding edge is incrementally enhanced.

[0016] As a further improvement to this technical solution, the hidden key point mining module includes: a meta-path definition unit, used to preset meta-path templates of various semantic types, wherein the meta-path template represents a multi-hop path pattern from the starting node to the target node.

[0017] The random walk sampling unit is used to perform random walk sampling on the heterogeneous information network according to the meta-path template, generate a node sequence, and input the node sequence into a graph neural network for node vector representation learning.

[0018] The implicit association calculation unit is used to calculate the multi-hop similarity score between electronic document nodes and known classified entity nodes based on the learned node vector representation. When the similarity score exceeds a preset similarity score threshold, it is determined that the electronic document has implicit classified points.

[0019] The counterfactual reasoning verification unit is used to recalculate the similarity score between the removed file node and the classified entity node by removing specific content paragraphs from the electronic file, and calculate the critical contribution of each paragraph. If the critical contribution exceeds the preset contribution threshold, the content paragraph is determined to be a key classified point, and a corresponding classification basis description is generated based on the content of the paragraph.

[0020] As a further improvement to this technical solution, the adversarial evolution watermark embedding module includes: a watermark generation network, used to receive the original content features of the electronic document and the confidentiality information to be embedded, and generate the document content after watermark embedding.

[0021] The attack simulation network receives the file content output by the watermark generation network, simulates various possible attack operations, including cropping, compression, noise addition, screenshotting, printing and scanning, format conversion, and attempts to extract watermark information from the attacked file.

[0022] The adversarial training unit is used to alternately train the watermark generation network and the attack simulation network. The goal of the watermark generation network is to generate watermarks that can still be completely extracted even after being attacked by the attack simulation network. The goal of the attack simulation network is to continuously discover the weaknesses of the watermark generation network in generating watermarks.

[0023] The co-occurrence feature encoding unit is used to encode watermark information as a small perturbation to the feature space of electronic document content, so that the statistical feature distribution of watermark and document content is integrated.

[0024] The self-healing decoding unit is used to reconstruct the damaged watermark information from the remaining file content by utilizing the completion capability of the generative model when it is detected that the content of the electronic document has been partially damaged, and output the reconstructed security information.

[0025] As a further improvement to this technical solution, the stress enhancement protection module includes: a behavior graph construction unit, used to construct a behavior sequence graph from the user's historical operation behavior, wherein the behavior sequence graph uses operation events as nodes and event time sequence and logical association as edges.

[0026] The real-time behavior acquisition unit is used to collect the user's current operation behavior sequence in real time.

[0027] The graph similarity calculation unit is used to calculate the similarity score between the current operation behavior sequence graph and the user's historical behavior graph. When the similarity score is lower than a preset threshold, it is determined that there is a behavior deviation.

[0028] The stress response unit, connected to the graph similarity calculation unit, is used to send a stress enhancement command to the adversarial evolution watermark embedding module when a behavioral deviation is determined, thereby triggering a dynamic increase in the watermark embedding strength.

[0029] As a further improvement to this technical solution, the behavior sequence graph is specifically implemented by: collecting the user's historical operation logs and extracting operation events as nodes, including file opening, file editing, file copying, file printing, file sending, and window switching.

[0030] Directed edges are constructed based on the chronological order and logical connections of events to form a graph of user historical behavior.

[0031] Vector representation learning is performed on the user's historical behavior graph to obtain the feature vector of the user behavior pattern, and the current operation behavior sequence collected in real time is constructed into the current behavior graph.

[0032] As a further improvement to this technical solution, the stress response unit is also used to: calculate the difference between a preset threshold and a similarity score as the behavioral deviation.

[0033] When the deviation of the behavior exceeds the first threshold but is lower than the second threshold, the adversarial evolution watermark embedding module is triggered to increase the watermark embedding strength to a medium level and overlay a dynamic screen watermark containing user identity information on the file display interface.

[0034] When the deviation of the behavior exceeds the second threshold, the adversarial evolution watermark embedding module is triggered to increase the watermark embedding strength to the highest level, while freezing the current access session or blurring the displayed content of the file.

[0035] As a further improvement to this technical solution, it also includes: a stroboscopic coding tracing module, used to embed invisible stroboscopic coding into the screen display content when the stress enhancement protection module triggers the highest level of protection; the stroboscopic coding includes the current user identity, device information, and timestamp, used to trace the source of the leak afterward by analyzing the stroboscopic coding in the secretly taken photos or videos.

[0036] As a further improvement to this technical solution, the stroboscopic coding traceability module includes: a coding generation unit, used to generate a unique stroboscopic coding sequence based on the current user identity, device information, and timestamp. The stroboscopic coding adopts a high-frequency flickering mode that is invisible to the human eye and is embedded through minute brightness changes of screen pixels.

[0037] The synchronous embedding unit is used to synchronously superimpose the strobe code with the displayed content when the screen display content is refreshed, so as to ensure that each frame of the display carries traceable encoding information.

[0038] The decoding and extraction unit is used to analyze suspected leaked photos or videos after the fact. It extracts the strobe code through image processing algorithms to restore the user's identity, device information and timestamp at the time of the leak, forming a legal-grade chain of evidence.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] 1. By constructing a heterogeneous information network and a meta-path random walk algorithm, multi-hop implicit relationships between documents and classified entities are mined, overcoming the limitations of traditional keyword matching and enabling the discovery of documents that are "not directly mentioned but are substantially classified." The counterfactual reasoning verification unit accurately locates key classified points and automatically generates classification criteria by removing paragraphs and calculating contribution, making the classification process interpretable and verifiable.

[0041] 2. A generative adversarial watermarking network is constructed. Through continuous adversarial training, watermark information capable of resisting various attacks such as cropping, compression, screenshotting, printing, and scanning is generated. The watermark and the file content feature space are fused and coexisted, so even if the file is screenshotted or photographed and leaked, the security level and source can still be traced. The self-healing decoding unit can reconstruct the watermark when the file is partially damaged, achieving a deep binding of "content is watermark".

[0042] 3. User historical operations are constructed into a behavioral graph. Behavioral patterns are learned through graph neural networks, and current behavior is compared in real time to calculate deviation. A tiered response is implemented based on the degree of deviation—medium-level response enhances the watermark and overlays a screen watermark; high-level response freezes the session or obscures content. Simultaneously, a stroboscopic encoding tracing module embeds invisible encoding at the highest level of protection. Afterwards, the identity, device, and time information of the leaker can be extracted from photos / videos, forming a legally binding chain of evidence, achieving a leap from passive auditing to proactive defense. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the system structure connection of the present invention.

[0045] Figure 2 This is a schematic diagram of the heterogeneous information network construction module structure of the present invention.

[0046] Figure 3 This is a schematic diagram of the hidden dense point mining module structure of the present invention.

[0047] Figure 4 This is a schematic diagram of the structure and connection of the anti-evolution watermark embedding module of the present invention.

[0048] Figure 5 This is a schematic diagram of the stress-enhanced protection module structure of the present invention.

[0049] Figure 6 This is a schematic diagram of the structure and connection of the stroboscopic encoding traceability module of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example: Please refer to Figure 1 As shown, a document intelligent classification and classification control system is provided, including: a heterogeneous information network construction module, used to construct a dynamically updated heterogeneous information network from personnel entities, document entities, project entities, equipment entities and their relationships within the target organization; the heterogeneous information network contains multiple types of nodes and multiple types of edges, and uses a time-series decay factor to dynamically adjust the weights of the edges.

[0052] The time decay factor refers to the time decay function.

[0053] In one specific embodiment, the heterogeneous information network construction module includes: a multi-source entity access unit, used to access data from the personnel information system, document management system, project management system, and equipment monitoring system within the organization, and to extract personnel nodes, document nodes, project nodes, and equipment nodes.

[0054] The relationship extraction unit is used to extract various types of relationships between entities, including the writing relationship between documents and people, the citation relationship between documents, the participation relationship between people and projects, the printing / sending relationship between documents and devices, and the chronological relationship between events and time.

[0055] The dynamic weight update unit is used to assign an initial weight to each associated edge and periodically decay the weight of the edge according to the time decay function. At the same time, when a new business interaction is detected, the weight of the corresponding edge is incrementally enhanced.

[0056] The time decay function is a composite function that takes into account the differences in multiple relation types:

[0057]

[0058] in, For a moment Time node With nodes The weight of the edges between them. Indicates the total number of relation types (write, cite, participate, print, send, etc.). Indicates the first The baseline weight of class relationships, Indicates the first The decay rate of class relationships (printed relationships decay quickly, participating relationships decay slowly). Indicates the first The time interval between the most recent occurrence of the class relationship Indicates the first Frequency enhancement coefficient of class relationship Indicates the number of times within the most recent time window The number of times class relationships occur. , This represents any two nodes in a heterogeneous information network. Indicates a relation type index.

[0059] The nodes here , It covers all entity types in the entire heterogeneous information network, including people, documents, projects, equipment, etc.

[0060] The hidden key information mining module is used to perform multi-hop association reasoning on the heterogeneous information network based on the meta-path random walk algorithm to mine key information in the electronic document content that is not directly mentioned but has an implicit association with known classified entities. It also verifies the critical contribution of the key information to the document's classification level through counterfactual reasoning and generates classification suggestions and basis for classification.

[0061] In one specific embodiment, the hidden key point mining module includes: a meta-path definition unit, used to preset meta-path templates of various semantic types, wherein the meta-path templates represent multi-hop path patterns from the starting node to the target node.

[0062] The random walk sampling unit is used to perform random walk sampling on the heterogeneous information network according to the meta-path template, generate a node sequence, and input the node sequence into a graph neural network for node vector representation learning.

[0063] The implicit association calculation unit is used to calculate the multi-hop similarity score between electronic document nodes and known classified entity nodes based on the learned node vector representation. When the similarity score exceeds a preset similarity score threshold, it is determined that the electronic document has implicit classified points.

[0064]

[0065] in, Represents electronic document nodes With known classified entity nodes Multi-hop similarity score between them This indicates the total number of preset metapath templates in the metapath definition unit. Indicates the meta-path index. Represented as the first Preset weights for element paths, Indicates from document node Depart, along the first The sequence of node types constrained by the element path template reaches the classified entity node through a random walk. The probability of.

[0066] The counterfactual reasoning verification unit is used to recalculate the similarity score between the removed file node and the classified entity node by removing specific content paragraphs from the electronic file, and calculate the critical contribution of each paragraph. If the critical contribution exceeds the preset contribution threshold, the content paragraph is determined to be a key classified point, and a corresponding classification basis description is generated based on the content of the paragraph.

[0067]

[0068] in, Paragraph Percentage of contribution to document confidentiality Indicates the removal of a paragraph The following document With classified entities Similarity score, Document node and classified entity nodes These are two nodes in a heterogeneous information network, and the edge between them also follows the update rules of the dynamic weight function.

[0069] An adversarial evolution watermark embedding module is used to construct a generative adversarial watermarking network. The generative adversarial watermarking network includes a watermark generation network and an attack simulation network. Through continuous adversarial training of the two, watermark information coexists with the feature space of the electronic document content, and the encryption result is encoded and embedded into the watermark information.

[0070] In one specific embodiment, the adversarial evolution watermark embedding module includes: a watermark generation network, used to receive the original content features of the electronic document and the security information to be embedded, and generate the document content after watermark embedding. The watermark generation network adopts a generator architecture of generative adversarial network.

[0071] The attack simulation network receives the file content output by the watermark generation network, simulates various possible attack operations, including cropping, compression, noise addition, screenshotting, printing and scanning, format conversion, and attempts to extract watermark information from the attacked file.

[0072] The adversarial training unit is used to alternately train the watermark generation network and the attack simulation network. The goal of the watermark generation network is to generate watermarks that can still be completely extracted even after being attacked by the attack simulation network. The goal of the attack simulation network is to continuously discover weaknesses in the watermarks generated by the watermark generation network. After training, the watermarks generated by the watermark generation network are immune to unknown attacks.

[0073] The symbiotic feature encoding unit is used to encode watermark information as a tiny perturbation to the feature space of the electronic document content, thereby integrating the statistical feature distribution of the watermark and the document content. This achieves the symbiosis of watermark and content.

[0074] The self-healing decoding unit is used to reconstruct the damaged watermark information from the remaining file content by utilizing the completion capability of the generative model when it is detected that the content of the electronic document has been partially damaged, and output the reconstructed security information.

[0075] Generative models refer to deep learning models used for semantic completion of missing content, specifically generators in generative adversarial networks (GANs). These generators, after training, learn the statistical distribution of electronic document content. When document content is partially corrupted, they can infer and complete the missing parts based on the remaining content, restoring a complete document semantically consistent with the original content, thus providing usable input for the watermark extraction subunit. In a specific embodiment of this invention, the generative model can share an architecture with the watermark generation network or be trained independently, but both rely on the generative capabilities of the GAN to achieve content restoration. The watermark generation network supports multi-level strength presets.

[0076] The self-healing decoding unit includes a content completion subunit and a watermark extraction subunit. The content completion subunit performs semantic completion on the missing content based on a generative adversarial network, and the watermark extraction subunit extracts watermark information from the completed content.

[0077] The stress-enhanced protection module is used to collect user operation behavior data in real time and construct a user behavior map. When the deviation between the current operation behavior and the user's historical behavior map exceeds a preset threshold, the dynamic watermark enhancement mode of the adversarial evolution watermark embedding module is triggered to improve the embedding strength and robustness of the watermark.

[0078] In one specific embodiment, the stress enhancement protection module includes: a behavior graph construction unit, used to construct a behavior sequence graph from the user's historical operation behavior, wherein the behavior sequence graph uses operation events as nodes and event time sequence and logical association as edges.

[0079] The real-time behavior acquisition unit is used to collect the user's current operation behavior sequence in real time.

[0080] The graph similarity calculation unit is used to calculate the similarity score between the current operation behavior sequence graph and the user's historical behavior graph. When the similarity score is lower than a preset threshold, it is determined that there is a behavior deviation.

[0081] Calculate the cosine similarity between the feature vector of the current behavior graph and the feature vector of the historical behavior pattern, and use it as the similarity score.

[0082] In another embodiment, the graph similarity calculation unit uses a Graph Attention Network (GAT) to encode the behavioral graph, specifically implemented as follows:

[0083] 1. Node Feature Definition

[0084] Each operation event node eigenvectors It is composed of the following information:

[0085] Operation type encoding: One-hot encoding (one-hot encoding watermark generation network) is used for operation types such as file opening, file editing, file copying, file printing, file sending, and window switching, with a dimension of 6;

[0086] File security level: The security level (unclassified, secret, confidential, top secret) of the file involved in the current operation is mapped to a 4-dimensional embedding vector;

[0087] Operation timestamp encoding: Maps operation time (hours, minutes) to a periodic encoding dimension of 4;

[0088] Operation object identification: If the operation involves a specific file or device, its global node ID in the graph is mapped to an 8-dimensional embedding vector.

[0089] Final node feature dimensions .

[0090] 2. Graph Neural Network Structure

[0091] A two-layer GAT is used, and the calculation formula for each layer of GAT is as follows:

[0092]

[0093] in, For the number of attention heads, Indicates splicing, Represented as the first Layer The normalized attention coefficients are calculated from each attention head. For trainable weight matrix, For ELU activation function, Represents a node The set of neighboring nodes in the graph The edges between them also follow the update rule of the dynamic weight function. After two layers of GAT, each node obtains a 64-dimensional feature vector. .

[0094] 3. Graph-level embedding vector generation

[0095] The graph-level embedding vector of the entire behavioral graph is obtained by average pooling:

[0096]

[0097] in, This represents the set of nodes in the graph. The graph-level embedding vector has a dimension of 64.

[0098] 4. Similarity Calculation

[0099] The similarity score between the current behavior graph and the historical behavior graph is the cosine similarity of their graph-level embedding vectors:

[0100]

[0101] in, Represents the current behavior graph. This represents a historical behavior graph, with a similarity score ranging from [−1, 1]. The closer the score is to 1, the more consistent the behavior patterns are.

[0102] 5. Training process and loss function

[0103] Graph neural networks are trained using a contrastive learning approach, with the goal of making the behavioral graph embeddings of the same user at different time segments as similar as possible, and the behavioral graph embeddings of different users as far apart as possible.

[0104] Training data construction:

[0105] Positive sample pairs: behavioral profiles constructed from two different time windows (e.g., morning and afternoon) of the same user;

[0106] Negative sample pairs: Behavioral profiles constructed from different users within the same time window.

[0107] The loss function uses InfoNCE loss (temperature parameter) )

[0108]

[0109] in, For anchor point map, For positive sample spectra, For negative sample maps, This represents the number of negative samples.

[0110] Once trained, the graph neural network can be used to calculate the similarity score between the current behavior graph and the user's historical behavior graph in real time.

[0111] The preset threshold is set based on the statistical distribution of user historical behavior data, for example, by subtracting twice the standard deviation from the mean of historical similarity scores; the first threshold and the second threshold are set according to security policy requirements, for example, 0.5 and 0.8 respectively. All three thresholds satisfy the condition: preset threshold < first threshold < second threshold.

[0112] The stress response unit, connected to the graph similarity calculation unit, is used to send a stress enhancement command to the adversarial evolution watermark embedding module when a behavioral deviation is determined, thereby triggering a dynamic increase in the watermark embedding strength.

[0113] In one specific embodiment, the behavior sequence graph is implemented by collecting the user's historical operation logs and extracting operation events as nodes, including file opening, file editing, file copying, file printing, file sending, and window switching.

[0114] Directed edges are constructed based on the chronological order and logical connections of events to form a graph of user historical behavior.

[0115] A graph neural network is used to learn the vector representation of the user's historical behavior map to obtain the feature vector of the user behavior pattern, and the current operation behavior sequence collected in real time is used to construct the current behavior map.

[0116] In one specific embodiment, the stress response unit is further configured to: calculate the difference between a preset threshold and a similarity score as the behavioral deviation.

[0117] When the deviation of the behavior exceeds the first threshold but is lower than the second threshold, the adversarial evolution watermark embedding module is triggered to increase the watermark embedding strength to a medium level and overlay a dynamic screen watermark containing user identity information on the file display interface.

[0118] When the deviation of the behavior exceeds the second threshold, the adversarial evolution watermark embedding module is triggered to increase the watermark embedding strength to the highest level, while freezing the current access session or blurring the displayed content of the file.

[0119] In one specific embodiment, it further includes: a stroboscopic coding tracing module, used to embed invisible stroboscopic coding into the screen display content when the stress enhancement protection module triggers the highest level of protection; the stroboscopic coding includes the current user identity, device information, and timestamp, and is used to trace the source of the leak by analyzing the stroboscopic coding in the secretly taken photos or videos afterward.

[0120] In one specific embodiment, the stroboscopic coding traceability module includes: a coding generation unit, used to generate a unique stroboscopic coding sequence based on the current user identity, device information, and timestamp. The stroboscopic coding adopts a high-frequency flickering mode that is invisible to the human eye and is embedded through minute brightness changes of screen pixels.

[0121] The synchronous embedding unit is used to synchronously superimpose the strobe code with the displayed content when the screen display content is refreshed, so as to ensure that each frame of the display carries traceable encoding information.

[0122] The decoding and extraction unit is used to analyze suspected leaked photos or videos after the fact. It extracts the strobe code through image processing algorithms to restore the user's identity, device information and timestamp at the time of the leak, forming a legal-grade chain of evidence.

[0123] In another specific embodiment:

[0124] I. System Deployment and Initialization

[0125] A military research institute has deployed the intelligent document classification and security management system described in this invention. The system is pre-connected to various information systems within the institute.

[0126] Personnel Information System: Includes basic information, job titles, and security classifications of all employees in the institute;

[0127] Document management system: Stores research reports, project documents, meeting minutes, etc. from previous years;

[0128] Project Management System: Records information on ongoing projects, including confidentiality level, participants, and timeframe;

[0129] Equipment monitoring system: tracks the usage records of equipment such as printers, copiers, and scanners.

[0130] The system extracts entity data from the aforementioned system and constructs an initial heterogeneous information network through the multi-source entity access unit in the heterogeneous information network construction module. The network includes the following node types:

[0131] Personnel nodes: such as "Zhang San (Researcher, Classification: Confidential)" and "Li Si (Assistant, Classification: Secret)";

[0132] Document nodes: such as "Technical Solution for XX Model Radar (V2.1)" and "Antenna Array Test Report", etc.

[0133] Project milestones: such as "XX Model Radar Project (Classification: Confidential)" or "YY Pre-research Project (Classification: Secret)";

[0134] Device nodes: such as "Printer (Department: Radar Room)" and "Scanner".

[0135] The relation extraction unit extracts the relationships between entities from historical data, for example:

[0136] Zhang San wrote the "Technical Solution for XX Model Radar (V2.1)";

[0137] The "XX Model Radar Technical Solution (V2.1)" references the "Antenna Array Test Report";

[0138] Zhang San participated in the "XX Model Radar Project";

[0139] The "Antenna Array Test Report" was printed on the printer.

[0140] The dynamic weight update unit assigns an initial weight to each associated edge and sets decay rate parameters for different relationship types (e.g., slow decay for write relationships and fast decay for print relationships), while simultaneously monitoring new business interactions in real time to dynamically enhance the weights. At this point, the dynamically updated heterogeneous information network is complete.

[0141] II. New Document Generation and Intelligent Password Determination

[0142] Researcher Zhang San began writing a new technical report, "An Overview of New Phased Array Radar Technology." During the writing process, the system's backend analyzed the document content in real time.

[0143] Hidden Key Discovery Module Launched:

[0144] The metapath definition unit pre-sets several metapath templates, such as:

[0145] Metapath 1: Documents → People → Projects

[0146] Metapath 2: Documents → Documents → People → Projects

[0147] Metapath 3: Documents → Equipment → Personnel → Project

[0148] Preset The path is divided into segments with weights of [0.4, 0.3, 0.15, 0.1, 0.05].

[0149] The random walk sampling unit performs a random walk based on the current heterogeneous information network (considering dynamic weights) to generate a node sequence, which is then input into the graph neural network (GAT) for node vector representation learning.

[0150] Implicitly related computational units calculate document nodes (New report) and known classified entity nodes Multi-hop similarity scores between (e.g., "XX model radar project") The calculation considers all meta-paths and performs a weighted sum. Assume the result is... (Preset similarity score threshold 0.6) If the score exceeds the threshold, the document is determined to contain hidden secrets.

[0151] The counterfactual reasoning verification unit locates key paragraphs in the document. The document mainly contains four paragraphs:

[0152] Introduction to the Development Background of Radar Technology

[0153] Key technical parameters of phased array radar (including specific values)

[0154] Future Development Trends Outlook

[0155] References list

[0156] The system removes each paragraph sequentially and calculates the similarity score after removal:

[0157] Remove back, Contribution ;

[0158] Remove back, Contribution ;

[0159] Remove back, Contribution ;

[0160] Remove back, Contribution .

[0161] The preset contribution threshold is 20%, therefore only (The section on key technical parameters) was identified as a critical classified point. The system automatically generated a description of the classification basis: "Section 2 of the document describes in detail the beamforming algorithm parameters and element spacing data of the phased array radar. These parameters are highly consistent with the core technical indicators of the 'XX Model Radar Project,' and it is recommended that the classification be 'Confidential.'" This classification recommendation was sent to the person responsible for classification for review and confirmation.

[0162] III. Counter-evolutionary watermark embedding

[0163] After the person responsible for classification approves the document, it is officially classified as "Confidential". At this point, the anti-evolutionary watermark embedding module is activated:

[0164] The watermark generation network receives the original content features and classification information (classification level, person responsible for classification, timestamp, etc.) of a document and generates document content with embedded watermark. The watermark is integrated into the text feature space of the document in the form of minor perturbations (e.g., by adjusting the font outline, subtle changes in line spacing, or by using semantic redundancy in the text for steganography).

[0165] During the training phase, the attack simulation network performs various simulated attacks (such as screenshotting, printing and scanning, format conversion, and adding noise) on documents generated by the watermark generation network, and attempts to extract the watermark from the attacked documents. Through adversarial training, the watermark generation network and the attack simulation network continuously evolve, and eventually the watermark generated by the watermark generation network can resist a variety of known and unknown attacks.

[0166] The co-occurrence feature encoding unit ensures that the watermark and the statistical feature distribution of the document content are integrated, making it difficult for attackers to detect differences even when comparing the original document with the watermarked document.

[0167] The self-healing decoding unit serves as a backup mechanism: if parts of the document are cut or corrupted in the future, a generative model (sharing an architecture with the watermark generation network) can be used to complete the missing content, and then the watermark can be extracted from it. For example, if someone maliciously deletes a paragraph containing key parameters, the system can still reconstruct the complete document and extract the security classification information from the remaining parts using generative completion.

[0168] IV. Stress Enhancement Protection During Document Use

[0169] Zhang San needs to access this confidential document. The system records his historical behavior pattern:

[0170] Historical behavior graph: Zhang San's typical operation pattern is: open document → edit → save → close, occasionally print. Graph neural networks learn its feature vectors.

[0171] On that day, after Zhang San opened the document, the real-time behavior acquisition unit detected an abnormal operation sequence: after opening the document, he quickly copied multiple paragraphs and frequently switched windows to external email pages, and the operation speed was significantly faster than usual.

[0172] The graph similarity calculation unit calculates the cosine similarity between the current behavior graph and the historical graph, which is 0.2 (preset similarity threshold 0.5), and determines that there is a behavior deviation. The behavior deviation is further calculated to be 0.3.

[0173] The stress response unit has a preset first deviation threshold of 0.1 and a second deviation threshold of 0.4. Since 0.1 < 0.3 < 0.4, a medium-level response is triggered.

[0174] Send a command to the adversarial evolution watermark embedding module to increase the watermark embedding strength from medium to high (making the watermark more robust).

[0175] A dynamic screen watermark is overlaid on the document display interface, containing the current user "Zhang San" and the real-time time. The watermark is displayed semi-transparently on the document background.

[0176] After Zhang San noticed the watermark, he stopped his abnormal actions. The system detected that his subsequent behavior returned to normal, the similarity score improved, and the risk was eliminated. If the deviation exceeds 0.4 (such as continuously attempting to take screenshots), the system will trigger the highest level of protection: immediately freeze the session, blur the document content, and log an alert.

[0177] V. Source Tracing of Leaks and Strobe Coding

[0178] Assuming an extreme scenario: A malicious user, Li Si, attempts to steal document content by taking a picture of the screen with his mobile phone. Since the system has already triggered the highest level of protection (or even if not triggered, the system actively enters a game-theoretic mode upon detecting the photographing activity), the stroboscopic encoding tracing module is activated:

[0179] The encoding generation unit generates a unique stroboscopic encoding sequence based on the current user identity, device information, and timestamp. This sequence is embedded in the pixel brightness changes of each frame of the screen in a high-frequency flickering manner that is invisible to the human eye.

[0180] The synchronous embedding unit ensures that each frame of the displayed image carries encoded information.

[0181] When photos or videos taken by Li Si are investigated afterward, security personnel input the suspected leaked photos into the system. The decoding and extraction unit extracts the strobe code through image processing algorithms (such as analyzing the time series of brightness changes) to restore the user's identity (Li Si), device information (terminal IP), and precise timestamp at the time of the leak, forming an undeniable chain of evidence.

[0182] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A document intelligent classification and classification control system, characterized in that, include: The heterogeneous information network construction module is used to construct a dynamically updated heterogeneous information network from personnel entities, document entities, project entities, equipment entities and their relationships within the target organization; the heterogeneous information network includes multiple types of nodes and multiple types of edges, and the weights of the edges are dynamically adjusted using a time-series decay factor; The hidden key point mining module is used to perform multi-hop association reasoning on the heterogeneous information network, to mine key points in the content of electronic documents that are not directly mentioned but have implicit associations with known classified entities, and to verify the key contribution of the key points to the classification level of the document, and to generate classification suggestions and classification basis. The adversarial evolution watermark embedding module is used to construct a generative adversarial watermarking network. The generative adversarial watermarking network includes a watermark generation network and an attack simulation network. Through continuous adversarial training of the two, watermark information coexists with the feature space of the electronic document content, and the encryption result is encoded and embedded into the watermark information. The stress-enhanced protection module is used to collect user operation behavior data in real time and construct a user behavior map. When the deviation between the current operation behavior and the user's historical behavior map exceeds a preset threshold, the dynamic watermark enhancement mode of the adversarial evolution watermark embedding module is triggered to improve the embedding strength and robustness of the watermark.

2. The document intelligent classification and classification control system according to claim 1, characterized in that: The heterogeneous information network construction module includes: The multi-source entity access unit is used to access data from the organization's personnel information system, document management system, project management system, and equipment monitoring system, and to extract personnel nodes, document nodes, project nodes, and equipment nodes. The relationship extraction unit is used to extract various types of relationships between entities, including the writing relationship between documents and people, the citation relationship between documents, the participation relationship between people and projects, the printing / sending relationship between documents and devices, and the chronological relationship between events and time. The dynamic weight update unit is used to assign an initial weight to each associated edge and periodically decay the weight of the edge according to the time decay function. At the same time, when a new business interaction is detected, the weight of the corresponding edge is incrementally enhanced.

3. The document intelligent classification and classification control system according to claim 1, characterized in that: The hidden key point mining module includes: Meta-path definition unit is used to preset meta-path templates of various semantic types. The meta-path templates represent multi-hop path patterns from the starting node to the target node. The random walk sampling unit is used to perform random walk sampling according to the meta-path template on the heterogeneous information network, generate a node sequence, and input the node sequence into a graph neural network for node vector representation learning. The implicit association calculation unit is used to calculate the multi-hop similarity score between electronic document nodes and known classified entity nodes based on the learned node vector representation. When the similarity score exceeds a preset similarity score threshold, it is determined that the electronic document has implicit classified points. The counterfactual reasoning verification unit is used to recalculate the similarity score between the removed file node and the classified entity node by removing specific content paragraphs from the electronic file, and calculate the critical contribution of each paragraph. If the critical contribution exceeds the preset contribution threshold, the content paragraph is determined to be a key classified point, and a corresponding classification basis description is generated based on the content of the paragraph.

4. The document intelligent classification and classification control system according to claim 1, characterized in that: The anti-evolutionary watermark embedding module includes: A watermark generation network is used to receive the original content features of electronic documents and the confidential information to be embedded, and to generate the document content after watermarking. The attack simulation network is used to receive the file content output by the watermark generation network, simulate various possible attack operations, including cropping, compression, noise addition, screenshotting, printing and scanning, format conversion, and attempt to extract watermark information from the attacked file. The adversarial training unit is used to alternately train the watermark generation network and the attack simulation network. The goal of the watermark generation network is to generate watermarks that can still be completely extracted even after being attacked by the attack simulation network. The goal of the attack simulation network is to continuously discover the weaknesses of the watermark generation network in generating watermarks. The co-occurrence feature coding unit is used to encode watermark information as a small perturbation to the feature space of electronic document content, so that the statistical feature distribution of watermark and document content is integrated. The self-healing decoding unit is used to reconstruct the damaged watermark information from the remaining file content by utilizing the completion capability of the generative model when it is detected that the content of the electronic document has been partially damaged, and output the reconstructed security information.

5. The document intelligent classification and classification control system according to claim 1, characterized in that: The stress-enhanced protection module includes: A behavior graph construction unit is used to construct a behavior sequence graph from the user's historical operation behavior. The behavior sequence graph uses operation events as nodes and event time sequence and logical association as edges. The real-time behavior acquisition unit is used to collect the user's current operation behavior sequence in real time; The graph similarity calculation unit is used to calculate the similarity score between the current operation behavior sequence graph and the user's historical behavior graph. When the similarity score is lower than a preset threshold, it is determined that there is a behavior deviation. The stress response unit, connected to the graph similarity calculation unit, is used to send a stress enhancement command to the adversarial evolution watermark embedding module when a behavioral deviation is determined, thereby triggering a dynamic increase in the watermark embedding strength.

6. The document intelligent confidentiality classification and classification control system according to claim 5, characterized in that: The behavioral sequence map is specifically implemented as follows: Collect users' historical operation logs and extract operation events as nodes, including file opening, file editing, file copying, file printing, file sending, and window switching; Directed edges are constructed based on the chronological order and logical connections of events to form a user history behavior graph; Vector representation learning is performed on the user's historical behavior graph to obtain the feature vector of the user behavior pattern, and the current operation behavior sequence collected in real time is constructed into the current behavior graph.

7. The document intelligent classification and classification control system according to claim 5, characterized in that: The stress response unit is also used for: The difference between the preset threshold and the similarity score is calculated as the behavioral deviation. When the deviation of behavior exceeds the first threshold but is lower than the second threshold, the adversarial evolution watermark embedding module is triggered to increase the watermark embedding strength to a medium level and overlay a dynamic screen watermark containing user identity information on the file display interface. When the deviation of the behavior exceeds the second threshold, the adversarial evolution watermark embedding module is triggered to increase the watermark embedding strength to the highest level, while freezing the current access session or blurring the displayed content of the file.

8. The document intelligent classification and classification control system according to claim 1, characterized in that: Also includes: The stroboscopic coding traceability module is used to embed invisible stroboscopic coding into the screen display content when the stress enhancement protection module triggers the highest level of protection; The stroboscopic code contains the current user's identity, device information, and timestamp, which is used to trace the source of the leak by analyzing the stroboscopic code in the secretly taken photos or videos afterward.

9. The document intelligent classification and classification control system according to claim 8, characterized in that: The strobe coding tracing module includes: The encoding generation unit is used to generate a unique strobe code sequence based on the current user identity, device information, and timestamp. The strobe code adopts a high-frequency flicker mode that is invisible to the human eye and is embedded through minute brightness changes of screen pixels. The synchronous embedding unit is used to synchronously superimpose the strobe code with the displayed content when the screen display content is refreshed, so as to ensure that each frame of the display carries traceable encoding information. The decoding and extraction unit is used to analyze suspected leaked photos or videos after the fact. It extracts the strobe code through image processing algorithms to restore the user's identity, device information and timestamp at the time of the leak, forming a legal-grade chain of evidence.

Images