Open source information exposed surface blocking method and system based on large model adversarial deduction

Abstract

The invention relates to the technical field of network security, in particular to an open source information exposed surface blocking method and system based on large model adversarial deduction, and solves the problem that traditional compliance auditing cannot defend big data puzzle type logic leakage. Analyzing target information by using a large language model subjected to adversarial fine adjustment, and generating a potential leakage reasoning chain for describing an attacker to deduce a secret-related conclusion from the information in combination with historical public data; constructing a mixed knowledge graph fusing public and confidential information, injecting target information as a test node, and establishing a logic association edge according to semantic features; executing reverse chain type deduction from the tentative node to the secret-related node in the knowledge graph, and calculating a reasoning path and cost based on association strength; and finally, the risk level is evaluated according to the quantified reasoning cost, and hierarchical blocking or intelligent desensitization treatment is executed. The crossing from passive rule matching to active simulation attacker logic deduction is realized, and the hidden leakage risk caused by multi-source information association can be effectively identified and blocked.

Description

Technical Field

[0001] This invention relates to the field of network security technology, specifically to an open-source information exposure surface blocking method and system based on large-scale adversarial simulation. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The internet contains a vast amount of open-source information, such as public data, tender documents, approval announcements, and commercial advertisements. Although this open-source information itself has undergone anonymization, the "big data jigsaw puzzle" effect allows attackers to leverage fragmented information for cross-validation and logical reasoning. By analyzing the correlation between multiple sources and fragmented information, attackers can indirectly deduce sensitive content (e.g., analyzing the equipment / facilities topology and operational patterns of the tendering party from tender documents and commercial advertisements), posing a highly concealed risk of logical leakage.

[0004] To address this risk, traditional blocking methods typically rely on static rule-based detection and manual review mechanisms, essentially acting as a passive "filter." Its core logic is to match known sensitive patterns (such as keywords or specific file types). It lacks the attacker's perspective and cannot simulate how, after acquiring the information to be released, it actively correlates, logically links, and progressively infers from publicly available information already released by relevant organizations. Therefore, it cannot detect indirect, logically chain-based leakage paths such as "Information A + Information B → Deducing Secret C." Summary of the Invention

[0005] To address the technical problems mentioned above, this invention provides an open-source information exposure surface blocking method and system based on adversarial inference using a large model. By simulating the attacker's thinking through a large model, it automatically collects publicly available fragmented information, performs reverse logical inference in a knowledge graph, predicts whether new information will form a "puzzle" leakage chain with historical data after its release, and automatically blocks or intelligently desensitizes information based on the risk level, thus blocking reasoning-based intelligence theft at the source.

[0006] The first aspect of this invention provides an open-source information exposure surface blocking method based on large-scale model adversarial inference, comprising the following steps:

[0007] The target information to be reviewed is obtained, and the target information is parsed using a large language model that has been adversarially fine-tuned to generate a semantic feature vector containing at least one potential leakage inference chain; wherein, the potential leakage inference chain is used to describe the sequence of steps that logically deduce the classified conclusion from the target information in combination with historical public data.

[0008] Construct a hybrid knowledge graph containing public information nodes and classified information nodes, and temporarily inject target information as a test node into the hybrid knowledge graph. Establish logical connection edges between test nodes and existing nodes based on semantic feature vectors.

[0009] In the hybrid knowledge graph, a reverse chain inference is performed from the probing node to the classified information node, and at least one inference path and its corresponding inference cost are calculated; wherein, the inference cost is determined based on the association strength of the logical association edges in the path;

[0010] Based on the reasoning cost, the logical leakage risk level of the target information is quantitatively assessed, and corresponding blocking or desensitization operations are performed on the target information according to the risk level.

[0011] Furthermore, the target information is parsed using a large language model that has been adversarially fine-tuned, specifically as follows:

[0012] The text and image modal data in the target information are respectively input into the visual-language cross-modal model to calculate the relevance weights between image regions and text descriptions;

[0013] Key information with relevance weights exceeding a preset threshold is retained as association anchors to generate structured cross-modal semantic metadata;

[0014] Cross-modal semantic metadata is input into a large language model that has been adversarially fine-tuned to separate explicit sensitive features from implicit logical anchors.

[0015] Furthermore, a hybrid knowledge graph containing both publicly available information nodes and classified information nodes is constructed, specifically as follows:

[0016] The graph is defined as a heterogeneous multipart graph G(t)=(V,E,W(t)), where the node set V is divided into a public domain node subset V_public and a secret domain node subset V_secret.

[0017] Assign weights w_ij(t) to the edges e(v_i,v_j) in the graph and make them follow a time decay model based on the intelligence half-life: w_ij(t)=w_0·e^{-λ(t-t_c)};

[0018] Where w_0 is the initial weight, λ is the decay factor, t_c is the edge creation time, and t is the current time;

[0019] When the edge weight is lower than the activation threshold, it is marked as an inactive edge and will no longer participate in the regular path calculation.

[0020] Furthermore, logical connections between the trial nodes and existing nodes are established based on the semantic feature vectors, specifically as follows:

[0021] By using a contrastive learning pre-trained model, the feature vectors of the trial nodes are mapped to the same semantic vector space as the feature vectors of existing nodes in the graph.

[0022] Calculate the cosine similarity between the feature vectors of the trial node and each existing node;

[0023] If the cosine similarity exceeds the preset logical association threshold, a logical association edge is created between the corresponding nodes, and the similarity is used as the initial weight of the edge.

[0024] Furthermore, at least one inference path and its corresponding inference cost are calculated, specifically as follows:

[0025] For each probe node, perform a search for the shortest logical reasoning path from that node to any node containing classified information in the hybrid knowledge graph;

[0026] The inference cost Cost(P) of path P is defined as the sum of the reciprocals of the weights of all edges on the path: Cost(P) = Σ_{e∈P}(1 / w_e);

[0027] Where w_e is the current weight of edge e;

[0028] Minimum inference cost min(Cost(P)) and path hop count are used as the core indicators for assessing the risk of logic leakage.

[0029] Furthermore, based on the risk level, corresponding blocking or desensitization measures are performed on the target information, specifically as follows:

[0030] If the risk level is set to automatic blocking, the system will automatically lock publishing permissions and trigger the circuit breaker mechanism.

[0031] If the risk level is intelligent desensitization level, a modification work order containing logical cutting suggestions will be automatically generated, and the semantic logic obfuscation engine or visual adversarial perturbation engine will be called to carry out collaborative defense processing of the target information.

[0032] Among them, the semantic logic obfuscation engine is used to identify and modify key reasoning anchors in the text, perform generalized descriptions, or inject fake noise data;

[0033] The intelligent visual desensitization engine is used to automatically locate and blur high-risk background areas in images based on relevance weights.

[0034] Furthermore, this also includes feedback evolution, specifically:

[0035] Regularly collect information on the countermeasures and feedback data on whether the red team has successfully cracked the system.

[0036] Based on feedback data, reinforcement learning algorithms are used to automatically adjust the weight parameters in the risk quantification function or update the step size parameters of the adversarial example generation strategy.

[0037] To achieve continuous dynamic optimization of risk assessment models and blocking strategies.

[0038] A second aspect of the present invention provides a system for implementing the above-described method, comprising:

[0039] The intelligent mapping and data acquisition module is configured to: simulate the intelligence gathering behavior of attackers, perform multimodal collection, anchor cleaning and entity mapping on the target unit's existing Internet exposure surface and internal data to be released, and provide the target information to be reviewed and the associated historical public data for the "Leakage Prevention-Oriented Related Semantic Analysis and Attack Surface Feature Extraction Module";

[0040] The semantic parsing and attack surface feature extraction module for leakage prevention is configured to: acquire the target information to be reviewed, and use a large language model that has been adversarially fine-tuned to parse the target information to generate a semantic feature vector containing at least one potential leakage inference chain; wherein, the potential leakage inference chain is used to describe the sequence of steps that logically deduce the classified conclusion from the target information and combined with historical public data.

[0041] The exposure risk mining module based on knowledge graph and reverse chain inference is configured to: construct a hybrid knowledge graph containing public information nodes and classified information nodes, temporarily inject target information as a test node into the hybrid knowledge graph, and establish logical connection edges between test nodes and existing nodes based on semantic feature vectors.

[0042] The exposure risk mining module based on knowledge graph and reverse chain inference is also configured to: perform reverse chain inference from the probing node to the classified information node in the hybrid knowledge graph, and calculate at least one inference path and its corresponding inference cost; wherein, the inference cost is determined based on the association strength of the logical association edges in the path.

[0043] The exposure risk assessment and intelligent blocking module is configured to: quantitatively assess the logical leakage risk level of the target information based on the inference cost, and perform corresponding blocking or desensitization operations on the target information according to the risk level.

[0044] A third aspect of the present invention provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the above-described open-source information exposure surface blocking method based on large-model adversarial inference.

[0045] A fourth aspect of the present invention provides an electronic device including at least one processor and a memory connected to the processor, the memory being used to store a computer program; the processor being used to execute the computer program, enabling the electronic device to implement the above-described open-source information exposure surface blocking method based on large-model adversarial inference.

[0046] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects:

[0047] 1. By employing an "adversarially fine-tuned large language model" and a "reverse chain deduction" mechanism, the system simulates the perspective of an attacker with advanced intelligence analysis capabilities. It not only understands the deep semantics of text but also proactively generates an "attack reasoning chain" that starts from the information to be released and incorporates historical internet data to make multi-step logical leaps. This allows the defender to "see" in advance how information might be used and interpreted after its release, thereby identifying and blocking seemingly harmless but actually crucial high-risk content at the source, advancing the defense from the information surface to the logical reasoning layer.

[0048] 2. Integrating a visual-language cross-modal model and unified semantic space alignment technology, it can calculate the strong semantic correlation between specific regions (such as equipment) in an image and text descriptions, automatically stripping and structuring these "association anchors." For example, it can identify the implicit correspondence between a specific model of equipment in the background of a typical engineering site image and the technical specifications mentioned in the text report. It can also detect and alert on complex risk scenarios where security relies on a single modality for auditing, but the combination of multiple modalities exposes core secrets.

[0049] 3. The weights of the edges in the knowledge graph automatically decay over time, ensuring that risk assessments are always based on current and effective threat exposure, rather than outdated information. By continuously collecting feedback on the actual effects of adversarial responses, reinforcement learning is used to automatically adjust the parameters of the risk assessment model and the adversarial generation strategy, achieving a shift from "manual upgrades" to "self-evolution." This allows the system to learn from each attack-defense interaction, continuously optimizing its ability to counter new inference attacks and puzzle-solving methods, thereby maintaining continuous protective effectiveness in highly dynamic adversarial environments.

[0050] 4. It provides a refined response process based on quantitative risk grading (automatic blocking, intelligent desensitization, and compliant release). Its core lies in the "semantic logic obfuscation engine" and the "intelligent visual desensitization engine." For text, the engine can locate and rewrite key anchor words and phrases that lead to logical reasoning, replacing specific references with generalized expressions, or injecting controllable noise into numerical values, thereby "severing" the reasoning chain while preserving the overall semantics and usability. For images, it can accurately identify and blur only areas containing sensitive background information, retaining the main content, achieving an optimal balance between security and utility, and ensuring the smooth progress of business disclosure. Attached Figure Description

[0051] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0052] Figure 1 A flowchart of a sensitive information exposure surface blocking method based on large-model adversarial inference provided for one or more embodiments of the present invention;

[0053] Figure 2 This is a schematic diagram of the architecture of a sensitive information exposure surface blocking system based on large-model adversarial inference, provided for one or more embodiments of the present invention. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0055] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0056] As described in the background section, traditional blocking methods are typically based on static detection using rule bases and manual review mechanisms. Essentially, they are passive "filters," whose core logic is matching known sensitive patterns (such as keywords or specific file types). They lack the attacker's perspective and cannot simulate how, after acquiring the information to be released, they actively correlate, logically link, and progressively infer it with publicly available information from relevant organizations. Therefore, they cannot discover indirect, logically chain-based leakage paths such as "Information A + Information B → Deducing Secret C." In other words, they lack adversarial reasoning capabilities and cannot predict the risk of logical leakage.

[0057] Meanwhile, traditional methods typically process data from different modalities, such as text and images, in isolation. For example, text moderation systems do not analyze accompanying images, and vice versa. However, advanced attacks excel at cross-modal associations (such as inferring the location of a project mentioned in text from a background building in an image). Existing technologies lack the ability to align and analyze textual semantics and visual features within a unified semantic space, thus failing to identify hidden leaks arising from "image-text comparisons." In other words, the detection dimension is singular, unable to handle association analysis involving multimodal data fusion.

[0058] Furthermore, traditional solutions rely on manual collection and analysis to update their rule bases and sensitive word lists, which is time-consuming and inefficient. Meanwhile, attackers' tools, data sources, and reasoning methods are constantly evolving. This speed difference between "static defense" and "dynamic attack" makes defense systems severely lagging and vulnerable to new or variant reasoning attacks, often only able to update rules after a data breach has occurred.

[0059] When a system (or a person) vaguely senses a risk, the traditional approach is often to completely block it or to simply and crudely cover it up (such as deleting the entire section or blurring the entire image). While this approach prevents potential leaks, it also severely undermines the business value and intended use of the information (such as for publicity or public disclosure), failing to strike a balance between preventing risks and preserving the value of the information.

[0060] Therefore, this solution provides an open-source information exposure surface blocking method and system based on adversarial inference using a large model. By simulating the attacker's thinking through a large model, it automatically collects publicly available fragmented information, performs reverse logical inference in a knowledge graph, predicts whether new information will form a "puzzle" leakage chain with historical data after its release, and automatically blocks or intelligently desensitizes information based on the risk level, thus blocking reasoning-based intelligence theft at the source.

[0061] This solution can simulate the perspective of an external attacker, and perform joint reverse reasoning based on the fragmented content to be released and historical public data to quantitatively assess whether there is a risk of core secrets being indirectly restored under the "big data jigsaw puzzle" effect; and achieve logical alignment of cross-modal data and in-depth mining of implicit relationship anchors, thereby constructing a proactive defense mechanism based on a large language model and hybrid knowledge graph, namely "footprint mapping-reverse reasoning-proactive blocking".

[0062] The framework consists of four main modules:

[0063] The intelligent mapping and data acquisition module adopts a reconnaissance task planner architecture driven by a large language model. It autonomously collects and arranges the potential exposure surface of open source intelligence sources associated with specific targets and the internal data to be released, and performs multimodal association anchor cleaning and cross-domain mapping of virtual identifiers to real entities.

[0064] The module for semantic parsing and attack surface feature extraction for leakage prevention utilizes a large language model that has been adversarially fine-tuned to separate explicit sensitive features and implicit logical anchors from the information to be disclosed, and simulates the generation of attack inference chains to extract structured risk feature vectors.

[0065] The exposure risk mining module based on knowledge graph and reverse chain inference is used to construct a hybrid heterogeneous graph that integrates historical public data and internal confidential data, and to mine the risk of logical leakage through cross-modal logical alignment and minimum inference cost search algorithm.

[0066] The exposure risk assessment and intelligent blocking module dynamically quantifies risks based on inference costs, executes automated hierarchical decisions, and calls upon semantic logic obfuscation and visual adversarial perturbation engines for collaborative defense.

[0067] The intelligent mapping and data acquisition module is used to simulate the intelligence gathering behavior of attackers and to map the target unit's existing internet exposure surface and internal data to be released.

[0068] This solution adopts the following... Figure 1 The steps shown achieve the blocking, specifically:

[0069] The system provides task instructions, formulates plans through a planner, collects multi-source data in a targeted manner, and performs streaming cleaning and hot data extraction to transform unstructured information into structured metadata with time-series attributes, providing data for risk mining.

[0070] By using association semantic analysis and thought chain reasoning, potential relationships between entities can be parsed from metadata, and quantifiable and computable risk feature vectors can be extracted to transform information into risk signals.

[0071] By injecting features into the knowledge graph and using a reverse chain inference algorithm, the shortest reasoning path from the information to be published to the confidential assets is explored, and the path reasoning cost is calculated, thereby quantifying the risk of the exposure surface.

[0072] Scoring and calibration are performed based on a dynamic risk quantification model, and tiered responses are triggered according to the risk level: automatic blocking, intelligent desensitization, or compliant release after adding watermarks, forming a complete closed loop from risk discovery to handling.

[0073] Combination Figure 2 This solution is described in detail. The intelligent mapping and data acquisition module adopts a strategy generation and task scheduling architecture driven by a large language model. Its core is the reconnaissance task planner and the "perception and execution layer" composed of multiple dedicated data processors.

[0074] Step 1: The reconnaissance task planner receives task instructions from higher levels that include entities, such as "Enter the name and project code of a unit to collect potential external exposure points related to it" or "Retrieve the complete set of engineering tender documents released by a unit." Based on a predefined task template library and rule library, and combined with the path analysis capabilities of a large language model, the planner generates one or more candidate data collection path schemes. Scheme evaluation mainly considers preset indicators such as coverage, relevance, and timeliness. The generated data collection scheme must be manually reviewed or approved by the rule engine before being sent to the execution layer. The execution layer calls the corresponding API or data collection tools according to the approved scheme to perform targeted data collection, avoiding infinite loops or uncontrollable behavior.

[0075] In this embodiment, the reconnaissance mission planner receives high-level instructions and a seed entity list. (For example: {"Project Code: X-732", "Supplier: Company A", "Person in Charge: Zhang San"}), and then a data collection path plan is provided. The final data collection path plan generated by the planner must be reviewed and confirmed by the administrator before it can be executed to ensure the controllability of the entire process.

[0076] Strategy planning and route expansion search: The data acquisition module does not perform a full network scan, but instead performs targeted searches based on predefined rules and templates.

[0077] against The initial search is performed by matching data from high-authority intelligence sources, such as bidding websites and specific channels on social media. If new related entities are found in the search results... (For example, if a new subcontractor B is mentioned in the tender document), the system will determine its value based on predefined association rules (such as entity type and co-occurrence frequency). If it meets the rules, then... Add to seed queue, set maximum number of associated hops (e.g., 3 jumps) to prevent the unordered spread of the search range through rule constraints.

[0078] The system uses Apache Flink to build a stream processing pipeline for the collected data and the data streams to be published internally. (Setting the sliding window size is also mentioned.) Seconds, using window functions Performs immediate filtering. (Function) Includes basic sensitive word filters, such as filtering out completely irrelevant advertising information, to ensure that only data streams with potential intelligence value proceed to the next stage.

[0079] Step Two: Identify anchor points that attackers can exploit. For image-text data, calculate the relevance weights between image regions (such as device nameplates or document headers) and text descriptions using a visual-language cross-modal model. Only retain weights exceeding the threshold. Key information: This process strips away irrelevant background, transforming unstructured data into structured metadata centered around the leak anchor point.

[0080] In this embodiment, this step is used to simulate the visual cognitive process of an attacker "finding puzzle pieces." The system deploys a large visual-language model based on the CLIP architecture. When the system inputs a site image of an engineering acceptance project to be published, the model not only identifies people but also focuses on non-moving features in the background, i.e., anchor points. Through attention weighting, the model calculates image regions, such as a specific entity region. With text description Relevance weights .

[0081] Set threshold Only when When the system determines that the feature is strongly correlated with the location description, it is considered a high-risk leakage feature.

[0082] According to the formula Output structured intelligence metadata, such as:

[0083] {Type:"Image_Anchor",Feature:"Server_Rack_Layout_Vector_X",Related_Text:"Unit_732_Core_System", Confidence: 0.88}. This constitutes the leak anchor point required for subsequent inference.

[0084] Step 3: Cross-domain identity mapping between virtual identifiers and real entities. The system identifies the collected virtual identifiers (such as project filing codes, official public account IDs, etc.) as graph nodes, and performs cross-validation using metadata (Exif) in the image, prominent company logos in the background, and uniform styles to confirm the specific business unit or system node from which the data originated, generating a trusted identifier. The entity attribution identifier (Lat, Lon).

[0085] In this embodiment, the system analyzes the officially published project number and external service account (virtual ID) collected, and combines them with the published project progress images, geographic mapping data in the tender attachments, and infrastructure features extracted in step two to perform multi-dimensional cross-validation. If all three point to the same area, a high-confidence value is generated. internal entities And create an edge in the graph that is (project number) --[located at]--> (internal entity).

[0086] A time decay function is introduced to dynamically manage the "timeliness risk" of intelligence. Each exposure mapping relationship is associated with a confidence score that decays over time. .in, As the initial confidence level, In order to gain time, This is the attenuation coefficient (which can be set for different data source types). The system performs periodic checks, when... At that time, the risk of direct leakage of the historical data is reduced, but it is still retained in the graph as historical background data (Context) for long-term correlation inference, to ensure that the graph truly reflects the current Internet exposure situation.

[0087] In this embodiment, a time decay function is applied to each mapping relationship. For example, a piece of information about "project implementation" was published 6 months ago ( (relatively large), its confidence level It will decrease significantly because it may have been abandoned. When Drop to threshold In the following cases, the node is not deleted, but instead marked as Historical_Context (old information). It will not trigger a high-risk alert on its own during subsequent reasoning, but will be reactivated if it forms a new chain of evidence with the latest data.

[0088] The semantic parsing and attack surface feature extraction module for information leakage prevention adopts a three-layer parsing system: "red team knowledge injection - logical leakage reasoning - cross-source association enhancement." Its core is a large language model that has undergone adversarial fine-tuning, and it integrates attack chain simulation and real-time open-source intelligence retrieval mechanisms. This enables the stripping and risk mapping of "explicit sensitive features" and "implicit logical anchors" in information to be released. Deployed on a high-performance computing server cluster, this module aims to use deep learning technology to identify the implicit logical leakage risks in information to be released.

[0089] Step 1: Efficient Construction and Knowledge Injection of the Red Team Adversarial Model. Based on a pre-trained large model using the Transformer architecture, the model is given an "intelligence gathering" perspective and catastrophic forgetting is avoided. Low-rank adaptation (LoRA) is used for fine-tuning. Essentially, this involves adjusting the parameters of the original model... By injecting a low-rank decomposition adapter, the forward propagation process becomes: .in, , , and rank Training only Two small matrices are all that's needed to quickly adapt the model to the open-source intelligence analysis (OSINT) field.

[0090] In this embodiment, a red team model capable of intelligence analysis and seeing through deception is trained. An open-source large language model with strong reasoning capabilities (such as DeepSeek-70B) is selected as the base Transformer model.

[0091] To avoid the high costs and catastrophic omissions associated with full fine-tuning, this embodiment uses the PEFT (Parameter-Efficient Fine-Tuning) library to implement LoRA fine-tuning.

[0092] Parameters are set to: fixed pre-trained weights Set the rank of the low-rank matrix scaling factor .

[0093] In the forward propagation, only calculations and updates are performed. and The two smaller matrices are calculated using the following formula: This allows the model to quickly adapt to intelligence analysis tasks while retaining general language comprehension capabilities.

[0094] Fine-tuning data It consists of a triple: (key asset list, business intelligence deduction chain, generalized description-specific model mapping table). Through this process, the model obtains a domain-specific "de-masking" mapping function. Generalized desensitized terms → classified entity vectors. For example, automatically mapping 'industrial-grade heavy water filtration components' to 'a specific type of nuclear power plant cooling system'; or mapping 'high-power anti-interference antenna' to 'cross-border emergency communication hub'.

[0095] In this embodiment, a dedicated dataset containing 500,000 samples is constructed. Sample example:

[0096] The proposed text states: "This project requires the procurement of special high-frequency signal generating components that are resistant to low temperatures and strong winds."

[0097] The inference chain states: "High-frequency signal acquisition components are typically used for cross-border high-bandwidth wireless communication or power grid monitoring; combined with recently released information on the construction of cross-border logistics hubs in the region," the true intent / entity is "the core design parameters or operating status of a certain cross-border high-load data link."

[0098] By training on this dataset, the model learned the mapping function. It is possible to directly associate general terms such as "low temperature resistance" and "high frequency" with specific critical infrastructure operational risks.

[0099] Step 2: Context-aware implicit logic leakage decoding and global semantic association modeling: Utilizing the Transformer's multi-head self-attention mechanism, the model computes any two tokens in the input sequence. , Correlation weights between : , Through this mechanism, the model can capture the strong association between 'frequency band signal tester' and contextual terms such as 'border trade zone' and 'unmanned operation', inferring that it actually refers to 'undisclosed cross-border drone logistics route testing', thereby inferring its... Does it exceed the risk threshold? .

[0100] In this embodiment, this step is performed during the inference phase, focusing on detecting the risk of "seemingly compliant but actually related." Given a proposed border patrol press release, the model uses a multi-head self-attention mechanism to calculate the relevance weights between lexical terms. Suppose that the phrases "special frequency band signal tester" and "special frequency band signal tester" appear in the text, spaced far apart. In conventional review, neither of these words would be considered sensitive. However, in the attention heatmap of this model, the weight between these two word units... Significantly higher than the threshold (e.g.) This indicates that the model recognizes a high probability that the combination of the two points points to "undisclosed core business data link stress testing." For the accompanying images, the system uses a ViT-L / 14 visual encoder to extract image features and maps them to the same embedding space as the text through a projection layer.

[0101] When the text mentions a location or includes an image, the model performs multimodal attention fusion to generate an enhanced semantic representation that incorporates environmental and visual contextual information. It is used to enhance the identification of the high-risk exposure triplet of "environmental characteristics-critical nodes-engineering stages" and to determine whether there is a risk of inferring the specific model and status of the node through background.

[0102] For example, the model calculates cross-attention between "time in the text (e.g., nighttime)" and "features in the image (e.g., noise features in low-light conditions)". If a very high match is found, an enhanced semantic representation is generated. This indicates that the combination of images and text poses a risk of exposing confidential information.

[0103] Step 3: Risk assessment based on attack chain simulation and dynamic retrieval. For ambiguous information intended for public release, the module does not perform simple compliance classification, but instead guides the model to generate a deductive logic chain of the attacker. Attempting to obtain information from public sources Derive the conclusion of confidentiality This process can be modeled as maximizing the attack success rate: If the model can generate a high-confidence inference chain, it proves that there is a risk of logical leakage of the information.

[0104] To simulate an attacker using publicly available historical data for cross-validation, the system first retrieves information from an external open-source intelligence repository during inference. Search for the most relevant information in the current context (including historical news and supplier directories). 1 record as background reference The final risk analysis results are derived from the model based on the original proposed release input. and external search fragments Co-generation: , This ensured that the assessment process considered the risk of data leakage resulting from piecing together old and new data.

[0105] In this embodiment, this step simulates an attacker using internet resources to piece together a puzzle. The system maintains a vector database based on FAISS or Milvus, storing open-source intelligence (OSINT) from the past 5 years, including past bidding records and supplier directories. Search process: When the input contains "supplier A", the system automatically retrieves the supplier's historical bidding records. .

[0106] The system inputs a prompt to the model: "You are a red team intelligence analyst. Given information X (currently planned for release) and background information R (search results), please try to deduce the true purpose of this project."

[0107] Model-generated thought chain. For example: "Supplier A specializes in producing high-performance AI computing chips (Background R), and the current project is procuring a large number of industrial-grade immersion liquid-cooled enclosures (Information X)." It is speculated that the facility is an undisclosed cross-border big data intelligent computing center. Conclusion: There is a risk of exposure to the critical infrastructure's computing power scale and business capacity limits. The system calculates the product of the generation probabilities of each logical jump in the inference chain as the "attack success rate". .like If so, then a logical leak is determined.

[0108] The final output of this module is a set of structured semantic feature vectors tailored for subsequent graph construction. in, For the parsed associated anchor points (such as equipment nameplates, document headers). The inherent sensitivity score, For inferable / reducible probabilities, This represents the confidence level of the model in its assessment of the risk.

[0109] The final output module outputs a structured feature vector. ,For example:

[0110] { "E_anchors": ["High-performance AI computing chip", "Supplier A", "Computing center"], / / Anchor points

[0111] "S_sensitivity": 0.4, / / Inherent sensitivity (low, because the words themselves are desensitized)

[0112] "P_inference": 0.92, / / Probability of inference (extremely high, complete logical chain)

[0113] "C_confidence": 0.95 / / Model confidence

[0114] }

[0115] The exposure surface risk mining module, based on knowledge graphs and reverse chain inference, adopts an architecture of "internal and external mapping graph structure and one-way attack simulation." The graph integrates "historical public data" and "internal high-level confidential data"; the core engine focuses on performing "reverse red team game simulation" to quantitatively assess the likelihood of data becoming an attack springboard before data release. This embodiment details the construction and operation process of the exposure surface risk mining module based on hybrid knowledge graphs and reverse chain inference. This module relies on high-performance graph databases (such as Neo4j or TigerGraph) and vector retrieval engines (such as Milvus) to build an internal sandbox environment for inference leakage testing before data release.

[0116] Step 1: Construction and maintenance of the "public-confidential" hybrid heterogeneous map.

[0117] The graph is defined as a heterogeneous multipart graph containing public and private domains. Among them, the node set Divide into two mutually exclusive core subsets: (Published nodes on the internet, such as past tender documents and news photos) and (Internal confidential nodes, such as equipment operation status). Edge The weight represents the logical deduction or co-occurrence relationship between the two. This indicates the credibility or correlation strength of the inference. Weight updates follow the intelligence half-life model: .in, This is the attenuation factor. For The node is set to "invisible" by the system, and is only visible when the red team simulates the logical distance, ensuring the security of the evaluation process itself.

[0118] In this embodiment, this step constructs a special graph environment that includes not only old data from the internet but also new data that may lead to data leaks. A heterogeneous multipart graph is constructed using the Neo4j graph database. .

[0119] Node partitioning and storage:

[0120] (Public domain nodes) Import historical data obtained from internet crawlers. For example, node Node_P1 (tag: News_Image, attributes: url=..., date=2023-01); node Node_P2 (tag: Bid_Doc, attribute: text="a certain project").

[0121] (Secret Domain Node): Imports an internal high-security database. For example, node Node_S1 (label: Secret_Base, attributes: asset_id='Core_Server_01', security_level='L4').

[0122] Stealth attribute implementation utilizes database access control lists (ACLs) or view mechanisms. The Secret tag node is invisible to the regular query interface; only the "red team attack simulation algorithm" is authorized to read the entire graph (RootPrivilege). Edge weights are dynamically updated (half-life mechanism), and a scheduled task (Cron Job) is deployed to execute the script daily at midnight. For each edge... Read its creation time and current time Applying the formula .set up (For general information) or (For important features). Pruning, when calculated... When this happens, the system marks the edge as Inactive, so that it no longer participates in the regular path calculation, in order to reduce the computational complexity of the graph.

[0123] Step 2: Constructing Cross-Modal Logical Associations Based on Vector Space Alignment. To simulate an attacker's ability to piece together intelligence from multiple data sources, implicit cross-modal association edges are established in the graph using the feature vectors output by the preceding module:

[0124] 1. Feature Space Mapping: Utilizing a contrastive learning pre-trained model, such as a fine-tuned version of CLIP, the features of the unstructured document to be published are mapped. Compared with existing image node features in the atlas Mapped to the same semantic vector space ;

[0125] 2. Logical edge generation: Calculate cosine similarity ,like If the attacker is deemed capable of associating the document with the image, the system automatically creates a "logical association edge" between the two nodes, using similarity as the initial edge weight. This simulates the behavior of attackers building intelligence puzzles.

[0126] In this embodiment, the CLIP model, fine-tuned by the preceding module, is used to integrate all image nodes in the atlas. Pre-encoded and stored in the Milvus vector library. When there are document nodes to be published. During input, its text feature vectors are converted into... .

[0127] Perform an approximate nearest neighbor search (ANN Search) in Milvus to find the nearest neighbor with... The top-K image nodes with the highest cosine similarity. A logical association threshold is set. .

[0128] Example: The document to be published mentions "a special insulating and anti-icing coating suitable for high-altitude permafrost environments." The image database contains a photo of a groundbreaking ceremony for a cross-border power transmission project published two years ago. Although the text of the two documents does not directly overlap, in the CLIP vector space... Operation: The system automatically creates an edge of type INFERRED_LINK between the "Anti-icing Coating Procurement Document" node and the "Transmission Project Ceremony Photo" node, with an initial weight of... .

[0129] The above steps simulate the process by which an attacker deduces that "this batch of special coatings was procured for winter maintenance of this specific cross-border transmission line".

[0130] Step 3: Probing Injection and Attack Surface Assessment. Simulating the attacker's perspective, execute a fully automated "probing-searching-quantification" process:

[0131] 1. Probing subgraph injection, specifically: using the information to be reviewed and published (anonymized text or images) as a set of probe nodes. Temporary injection into the current graph And based on step two, it generates its connection with neighboring nodes in real time;

[0132] 2. Core secret reachability query: The system does not traverse the entire graph, but instead performs a query on each probe node. Calculate its distance to any core security node The shortest logical inference path. This is transformed into a weighted shortest path search problem on a graph (such as using Dijkstra's or A* algorithms).

[0133] In this embodiment, a temporary transaction in the graph database is initiated through trial subgraph injection (Sandbox Injection), or an in-memory graph of Neo4j GDS is used to test the information to be released (such as a de-identified "cafeteria procurement list") as a trial node. Insert the memory graph and connect it to the existing one based on the logical edges generated in step two. Nodes. Note: This operation does not write to disk, is rolled back after evaluation, and does not pollute the main graph. An improved Dijkstra's algorithm or Yen's K-Shortest Paths algorithm is used to find nodes. To any The path.

[0134] Define path Reasoning cost (Higher edge weights indicate stronger associations and lower inference costs). If there exists a path such that... (i.e., there exists a shortcut that is very easy to deduce), or the number of path jumps. If the information is reachable within 3 hops, it is determined that the information to be published poses a high risk of logical leakage. The system outputs the path with the highest risk value. The document provides a complete chain of evidence showing how the "information to be released" ultimately led to the leakage of confidential information through public disclosure, thus providing a target for subsequent intervention.

[0135] In this embodiment, the path cost function , border rights The closer the edge weight is to 1 (the stronger the correlation), the closer the cost is to 1 (the easier the reasoning); the smaller the edge weight, the greater the cost tends to be.

[0136] The threshold for logical shortcuts is set as follows: If the total cost of a certain path (This means that only about 3 steps of strong association reasoning are needed) to be classified as high-risk. Jump count threshold: .like This is considered direct exposure. Example: Path: Cafeteria Procurement List (probing) -- (Relevance 0.9) --> Supplier Delivery Records (Public) -- (Relevance 0.8) --> Some Hidden Internal Project Information (Confidential). Cost: .in conclusion: This triggered a red alert.

[0137] The above path The serialized output as a JSON object explicitly states: "Although the purchase list itself only states delivery to point A, combined with the publicly available supplier records for point A (which is also point B) on the internet, it can be directly deduced to confidential information B." This provides a solid chain of evidence for subsequent blocking.

[0138] The Exposure Surface Risk Assessment and Intelligent Blocking module addresses the core issues of traditional compliance systems: "single-point compliance but overall data leakage" and "lack of targeted remediation recommendations." By constructing an autonomous system integrating data leakage risk quantification, automated hierarchical decision-making, multimodal adversarial defense, and performance feedback, it achieves a closed loop from discovering inference vulnerabilities to severing logical links. This module operates according to a clear four-stage autonomous cycle: risk quantification → hierarchical decision-making → adversarial response → feedback evolution. As an intelligent gateway for data dissemination, this module is deployed on a control server at the boundary of internal and external network exchanges, integrating a probability calculation engine and a multimodal adversarial sample generator.

[0139] Step 1: Dynamic Risk Quantification Based on Inference Cost. This stage integrates the reverse inference path output from the preceding modules with the associated anchor points into a statistically significant unified leakage risk indicator. For each unit of information to be released, a feature vector is generated. ,in, Based on the inherent sensitivity of the information itself, and the definition of security classification; The probability of successfully constructing an inference chain for the red team model. The factors that led to the exposure of core secrets after the leak. To assess the freshness of historical publicly available data, a weighted fusion function is defined, where the inference probability is related to the freshness of the data. Assigning the highest weight yields the original risk value: .in This is the time-correlation enhancement factor. To avoid score drift, it is input into the temperature scaling calibrator. Output a final leakage risk score with a good probabilistic interpretation: This calibration process ensures It can accurately reflect the probability of information being "reconstructed" after it is released.

[0140] In this embodiment, this step transforms qualitative logical links into quantitative probabilities. For each information unit entering the review process... The system automatically generates four-dimensional feature vectors. , (Inherent sensitivity) Query the internal security classification table, such as "Ordinary News" = 0.1, "Device Parameters" = 0.6. Attack success rate. As the probability of inference . The (impact factor) is calculated based on graph centrality. If the exposed point connects to edge nodes, If connected to a core security node, . (Related Freshness) By querying the most recent update time of historical publicly available data. (Day), Calculation Applying the weighted fusion formula: Set weights , The (inference probability) factor has the highest weight, reflecting the core idea of ​​"anti-correlation." To address the issue of overly concentrated or divergent model output probability distributions, a variant of the Sigmoid algorithm is used as a calibrator. Among them, temperature parameters and deviation The optimization is achieved by minimizing the negative log-likelihood loss (NLL) on the validation set (historical leak / non-leak samples). Final output. .

[0141] Step Two: Automated hierarchical decision-making based on risk scores, with a three-level response mechanism as follows:

[0142] Automatic blocking level If information is deemed high-risk due to "direct" leakage logic (such as text content that directly exposes core information), the system will automatically lock the publishing permissions and trigger the circuit breaker mechanism.

[0143] Intelligent desensitization level If information is deemed to have a "transparent connection" risk, the system will automatically generate a modification request containing "logical disconnection suggestions" and "adversarial examples" and push it to the review panel.

[0144] Compliance release level If the data is deemed logically secure, it will be automatically watermarked and tracked with a digital watermark before being published and included in the Internet footprint monitoring queue.

[0145] In this embodiment, dual thresholds are set based on historical system operation data. (High-risk circuit breakers) and (Desensitization threshold). The decision-making logic is: Case A ( ): The system triggers the circuit breaker mechanism. The API interface directly returns a 403 Forbidden error, locks the data publishing permission, and sends a "Level 1 Leakage Blocking Notice" to the security administrator, along with a complete logical chain of evidence. Case B ( The system generates a modification work order. The data is suspended and enters the "Pending De-identification" queue, along with suggestions generated in step three (such as "suggestion to obfuscate the description"). Case C ( The system executes compliant release. It calls a digital watermarking algorithm (such as DCT frequency domain watermarking) to embed the publisher's ID and timestamp into the material, then pushes it to the public gateway and notifies Module 1 to include it in the subsequent "Internet footprint monitoring" scope.

[0146] Step 3: Collaborative intelligent countermeasures and interventions: Based on the data modality and leakage path type, different defense engines are invoked to perform interventions.

[0147] Semantic logic obfuscation engine (for text / data leaks): For text that triggers intelligent de-identification, it calls the "logic truncation" submodule based on a domain-wide model. This submodule not only performs traditional keyword replacement (such as "..."), but also... The core of this approach, which involves identifying and modifying "key inference anchors," is to identify and modify specific inference anchors. For example, it can automatically generalize "specific model V5.0 test" to "next-generation system test" or inject false noise data to break the attacker's inference chain or lead to incorrect conclusions while preserving the business semantics.

[0148] The intelligent visual desensitization engine (for image leakage) aims to remove sensitive visual anchors in the background that may lead to logical deductions while preserving the business value of the main image subject. Algorithm implementation: Based on high-weighted associated anchors identified in the preceding data acquisition and feature extraction steps, or by recalculating the semantic relevance weights between image features and sensitive text descriptions, it automatically locates high-risk background areas containing clues to leakage. Gaussian blur or mosaic masking is automatically applied to anchor areas where the relevance weight exceeds a threshold.

[0149] In this embodiment, the semantic logic obfuscation engine does not perform simple "asterisk replacement," but rather rewrites the entire model. Input: The intended text to be published is "...a full-load mounting test of a cross-border microwave relay node was conducted at a border pass...". Combining this with the preceding module's identification of a high-risk inference chain consisting of "specific pass name" + "microwave relay," a "generalization strategy" is selected. Output rewritten: "Performance verification of critical communication infrastructure was conducted under complex weather conditions...". For non-critical values ​​(such as the quantity of materials), obfuscation (noise injection) is applied, preventing attackers from accurately inferring the node's actual service capacity through material and energy consumption throughput, while maintaining data availability at the order-of-magnitude level.

[0150] The intelligent visual desensitization engine replaces traditional full-image masking or physical decryption methods. It aims to preserve the business value of the image subject (such as project progress or compliant equipment) while removing sensitive visual anchors in the background that could lead to logical deductions. It reuses the image region-text relevance weights calculated based on a large vision-language model from the preceding data acquisition module. and the corresponding image region coordinates Alternatively, during the blocking phase, the CLIP model can be invoked again to calculate the cosine similarity between each image patch and the sensitive intent description, thereby automatically locating high-risk areas.

[0151] For correlation weights exceeding the high-risk threshold ( For discrete regions (such as equipment nameplates or specific landmark buildings), Gaussian blur or mosaic masking is automatically applied to prevent their visual features from being recognized by the feature extractor.

[0152] The feedback evolution (closed-loop) system periodically collects feedback data after processing, such as whether the adversarial processed information was successfully cracked by the subsequent red team test. If a certain obfuscation strategy is found to be ineffective, the system uses reinforcement learning (RL) to automatically adjust the weights of the risk quantification function. Alternatively, the step size parameter for generating adversarial examples can be updated to achieve continuous evolution of defensive capabilities.

[0153] Traditional technologies rely on static rule bases, essentially employing a "wait-and-see" defense approach based on historically known patterns. This approach is completely ineffective against new threats where attackers proactively piece together fragmented public information. This invention systematically simulates the perspective of an attacker with advanced intelligence analysis capabilities through an "adversarially fine-tuned large language model" and a "reverse chain deduction" mechanism. This model not only understands the deep semantics of text but also proactively generates an "attack reasoning chain" that starts from the information to be released and incorporates historical internet data to make multi-step logical leaps. This allows the defender to "see" in advance how information might be used and interpreted after its release, thereby identifying and blocking seemingly harmless but actually crucial high-risk content at the source, advancing the defense from the information surface to the logical reasoning layer.

[0154] Existing auditing systems often analyze different modalities of information, such as text, images, and data, in isolation. Advanced attacks, however, excel at cross-verifying the truth from cross-modal clues like "inconsistencies between text and images" and "background hints." To address this deficiency, this solution integrates a visual-language cross-modal model and a unified semantic space alignment technique. It can calculate the strong semantic correlation between specific regions in an image (such as equipment or landmarks) and text descriptions, automatically stripping and structuring these "linking anchors." For example, it can identify the implicit correspondence between a specific model of equipment in the background of a typical engineering site image and the technical specifications mentioned in the text report. This capability enables the system to discover and alert on complex risk scenarios where relying on a single modal audit is absolutely secure, but combining multiple modalities exposes core secrets.

[0155] Traditional defense systems rely on manual case analysis and rule base maintenance for updates, resulting in long cycles, slow responses, and perpetual lag behind the evolution of attack techniques. This solution introduces a dynamic "intelligence half-life" knowledge graph and a closed-loop feedback evolution mechanism based on reinforcement learning. The weights of the related edges in the knowledge graph automatically decay over time, ensuring that risk assessments are always based on the current effective threat exposure, rather than outdated information. More importantly, by continuously collecting feedback on the actual effects of adversarial responses, the system uses reinforcement learning to automatically adjust the parameters of the risk assessment model and the adversarial generation strategy, achieving a shift from "manual upgrades" to "self-evolution." This allows the system to learn from every attack and defense interaction, continuously optimizing its ability to counter new inference attacks and jigsaw puzzle methods, thereby maintaining continuous protective effectiveness in a highly dynamic adversarial environment.

[0156] Traditional security measures, when faced with ambiguous risks, often resort to crude methods like "simple interception" or "complete masking." While this avoids leaks, it severely undermines the legitimate uses of information, such as public disclosure and promotion. To address this contradiction, this solution provides a refined response process based on quantitative risk grading (automatic blocking, intelligent desensitization, and compliant release). Its core lies in a "semantic logic obfuscation engine" and an "intelligent visual desensitization engine." For text, the engine can locate and rewrite key anchor words and phrases that lead to logical reasoning, replacing specific references with generalized expressions or injecting controllable noise into numerical values, thereby "severing" the reasoning chain while preserving the overall semantics and usability. For images, it can accurately identify and blur only areas containing sensitive background information, retaining the main content. This precise handling achieves an optimal balance between security and utility, ensuring the smooth operation of public business.

[0157] Correspondingly, a computer program product includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the aforementioned open-source information exposure surface blocking method based on large-model adversarial inference.

[0158] Correspondingly, an electronic device includes at least one processor and a memory connected to the processor, the memory being used to store a computer program; the processor is used to execute the computer program, enabling the electronic device to implement the above-described open-source information exposure surface blocking method based on large-model adversarial inference.

[0159] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An open-source information exposure surface blocking method based on large-scale model adversarial inference, characterized in that, Includes the following steps: The target information to be reviewed is obtained, and the target information is parsed using a large language model that has been adversarially fine-tuned to generate a semantic feature vector containing at least one potential leakage inference chain; wherein, the potential leakage inference chain is used to describe the sequence of steps that logically deduce the classified conclusion from the target information in combination with historical public data. Construct a hybrid knowledge graph containing public information nodes and classified information nodes, and temporarily inject target information as a test node into the hybrid knowledge graph. Establish logical connection edges between test nodes and existing nodes based on semantic feature vectors. In the hybrid knowledge graph, a reverse chain inference is performed from the probing node to the classified information node, and at least one inference path and its corresponding inference cost are calculated; wherein, the inference cost is determined based on the association strength of the logical association edges in the path; Based on the reasoning cost, the logical leakage risk level of the target information is quantitatively assessed, and corresponding blocking or desensitization operations are performed on the target information according to the risk level.

2. The open-source information exposure surface blocking method based on large-model adversarial inference as described in claim 1, characterized in that, The target information is parsed using a large language model that has been adversarially fine-tuned, specifically as follows: The text and image modal data in the target information are respectively input into the visual-language cross-modal model to calculate the relevance weights between image regions and text descriptions; Key information with relevance weights exceeding a preset threshold is retained as association anchors to generate structured cross-modal semantic metadata; Cross-modal semantic metadata is input into a large language model that has been adversarially fine-tuned to separate explicit sensitive features from implicit logical anchors.

3. The open-source information exposure surface blocking method based on large-model adversarial inference as described in claim 1, characterized in that, Construct a hybrid knowledge graph that includes both publicly available information nodes and classified information nodes, specifically as follows: The graph is defined as a heterogeneous multipart graph G(t)=(V,E,W(t)), where the node set V is divided into a public domain node subset V_public and a secret domain node subset V_secret. Assign weights w_ij(t) to the edges e(v_i,v_j) in the graph and make them follow a time decay model based on the intelligence half-life: w_ij(t)=w_0·e^{-λ(t-t_c)}; Where w_0 is the initial weight, λ is the decay factor, t_c is the edge creation time, and t is the current time; When the edge weight is lower than the activation threshold, it is marked as an inactive edge and will no longer participate in the regular path calculation.

4. The open-source information exposure surface blocking method based on large-model adversarial inference as described in claim 1, characterized in that, Logical connections between test nodes and existing nodes are established based on semantic feature vectors, specifically as follows: By using a contrastive learning pre-trained model, the feature vectors of the trial nodes are mapped to the same semantic vector space as the feature vectors of existing nodes in the graph. Calculate the cosine similarity between the feature vectors of the trial node and each existing node; If the cosine similarity exceeds the preset logical association threshold, a logical association edge is created between the corresponding nodes, and the similarity is used as the initial weight of the edge.

5. The open-source information exposure surface blocking method based on large-model adversarial inference as described in claim 1, characterized in that, Calculate at least one inference path and its corresponding inference cost, specifically: For each probe node, perform a search for the shortest logical reasoning path from that node to any node containing classified information in the hybrid knowledge graph; The inference cost Cost(P) of path P is defined as the sum of the reciprocals of the weights of all edges on the path: Cost(P) = Σ_{e∈P}(1 / w_e); Where w_e is the current weight of edge e; Minimum inference cost min(Cost(P)) and path hop count are used as the core indicators for assessing the risk of logic leakage.

6. The open-source information exposure surface blocking method based on large-model adversarial inference as described in claim 1, characterized in that, Based on the risk level, corresponding blocking or desensitization measures are performed on the target information, specifically: If the risk level is set to automatic blocking, the system will automatically lock publishing permissions and trigger the circuit breaker mechanism. If the risk level is intelligent desensitization level, a modification work order containing logical cutting suggestions will be automatically generated, and the semantic logic obfuscation engine or visual adversarial perturbation engine will be called to carry out collaborative defense processing of the target information. Among them, the semantic logic obfuscation engine is used to identify and modify key reasoning anchors in the text, perform generalized descriptions, or inject fake noise data; The intelligent visual desensitization engine is used to automatically locate and blur high-risk background areas in images based on relevance weights.

7. The open-source information exposure surface blocking method based on large-model adversarial inference as described in claim 1, characterized in that, This also includes feedback evolution, specifically: Regularly collect information on the countermeasures and feedback data on whether the red team has successfully cracked the system. Based on feedback data, reinforcement learning algorithms are used to automatically adjust the weight parameters in the risk quantification function or update the step size parameters of the adversarial example generation strategy. To achieve continuous dynamic optimization of risk assessment models and blocking strategies.

8. An open-source information exposure surface blocking system based on large-scale model adversarial simulation, characterized in that: include: The intelligent mapping and data acquisition module is configured to: simulate the intelligence gathering behavior of attackers, perform multimodal collection, anchor cleaning and entity mapping on the target unit's existing Internet exposure surface and internal data to be released, and provide the target information to be reviewed and the associated historical public data for the "Leakage Prevention-Oriented Related Semantic Analysis and Attack Surface Feature Extraction Module"; The semantic parsing and attack surface feature extraction module for leakage prevention is configured to: acquire the target information to be reviewed, and use a large language model that has been adversarially fine-tuned to parse the target information to generate a semantic feature vector containing at least one potential leakage inference chain; wherein, the potential leakage inference chain is used to describe the sequence of steps that logically deduce the classified conclusion from the target information and combined with historical public data. The exposure risk mining module based on knowledge graph and reverse chain inference is configured to: construct a hybrid knowledge graph containing public information nodes and classified information nodes, temporarily inject target information as a test node into the hybrid knowledge graph, and establish logical connection edges between test nodes and existing nodes based on semantic feature vectors. The exposure risk mining module based on knowledge graph and reverse chain inference is also configured to: perform reverse chain inference from the probing node to the classified information node in the hybrid knowledge graph, and calculate at least one inference path and its corresponding inference cost; wherein, the inference cost is determined based on the association strength of the logical association edges in the path. The exposure risk assessment and intelligent blocking module is configured to: quantitatively assess the logical leakage risk level of the target information based on the inference cost, and perform corresponding blocking or desensitization operations on the target information according to the risk level.

9. A computer program product, characterized in that, Includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the open-source information exposure surface blocking method based on large-model adversarial inference as described in any one of claims 1-7.

10. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, the memory being used to store computer programs; the processor is used to execute the computer programs, enabling the electronic device to implement the open-source information exposure surface blocking method based on large model adversarial inference as described in any one of claims 1-7.

Images