Intelligent archive opening identification method based on large model

Abstract

The invention discloses an intelligent archive opening and identifying method based on a large model, particularly relates to the technical field of archive data auditing, and is used for solving the problems of insufficient cross-modal data analysis capability, lagging rule updating and low man-machine cooperation efficiency in the prior art. Fusing cross-modal features of texts, images and metadata through a hybrid expert model to generate multi-modal feature vectors, and dynamically allocating the multi-modal feature vectors to a rule network, a semantic network and a domain network for cooperative processing based on attention weights; the rule network parameters are optimized through gradient projection constraint, and regulation-driven real-time adaptation is achieved; matching sensitive data in combination with a multi-dimensional feature matrix of auditing personnel, and optimizing task allocation accuracy; removing redundant links by utilizing value flow analysis to generate a lightweight process, and recording as a tamper-proof evidence chain through a block chain evidence storage solidification operation; the auditing efficiency and accuracy are improved, and the compliance traceability is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of archive data auditing, and more specifically, to an intelligent archive open identification method based on a large model. Background Art

[0002] Archives access appraisal involves identifying sensitive information in multimodal archival data, assessing compliance, and conducting cross-departmental collaborative reviews. Existing technologies typically employ a review model that combines automated initial screening with manual review based on a rules engine. This review is accomplished through predefined keyword matching, text feature extraction, and the construction of a human experience database. Technically, existing methods rely on traditional machine learning models to classify and process structured data, and process management systems to assign tasks and provide feedback.

[0003] However, when processing unstructured archival data, existing technologies struggle to meet the accuracy and efficiency of open authentication due to limitations in the semantic understanding capabilities of models. Specifically, traditional models are unable to parse implicit semantic associations within multimodal archives, making it difficult to effectively identify cross-modal sensitive information. Furthermore, the lag in adapting audit rules to dynamic regulations, as well as the synergy gap between human experience and automated processes, further exacerbate the risk of misjudgment and resource consumption, hindering the efficiency of open authentication. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a large-scale model-based intelligent open archive identification method to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for intelligent open identification of archives based on a large model includes the following steps:

[0007] S1. Perform cross-modal semantic analysis on the text data, image data and metadata of the archives, and generate multimodal feature vectors based on the hybrid expert model;

[0008] S2. dynamically assigning the archival data to at least one processing path among a rule network for performing risk keyword matching, a semantic network for identifying cross-modal sensitive associations, and a domain network for matching regional features based on the attention weight distribution of each data modality in the multimodal feature vector;

[0009] S3. Based on the regulatory update directive, the weight parameter optimization method based on gradient projection constraints is used to locally adjust the weight parameters related to risk keyword matching in the rule network;

[0010] S4. Construct a multi-dimensional professional feature matrix for auditors, perform similarity matching between the cross-modal sensitive correlation data identified by the semantic network and the multi-dimensional professional feature matrix, and assign them to the audit nodes corresponding to the qualifications;

[0011] S5. Perform value stream analysis on the audit process assigned to the audit node, remove manual confirmation steps and paper transmission nodes, and generate a lightweight audit process;

[0012] S6. Write the multimodal feature vector, the dynamically allocated processing path, and the operation record of the lightweight audit process into the blockchain to generate a timestamped evidence chain.

[0013] In a preferred embodiment, S1 includes:

[0014] Perform cross-modal alignment on the text data, image data, and metadata of the archive, calculate the cross-modal semantic correlation between the text and image, and generate a cross-modal alignment feature vector;

[0015] The cross-modal aligned feature vectors are input into the text expert network, image expert network, and metadata expert network in the hybrid expert model to extract text deep semantic features, image spatial features, and metadata spatiotemporal features, respectively.

[0016] Perform multimodal feature fusion on text deep semantic features, image spatial features, and metadata spatiotemporal features, and calculate the weight coefficients of different modal features through a multi-head attention mechanism;

[0017] The text deep semantic features, image spatial features and metadata spatiotemporal features are weighted and summed based on the weight coefficients to generate a multimodal feature vector.

[0018] In a preferred embodiment, S2 includes:

[0019] Normalize the attention weight distribution of text, image and metadata in the multimodal feature vector to generate text weight value, image weight value and metadata weight value;

[0020] Generate at least one of a rule network trigger flag, a semantic network trigger flag, and a domain network trigger flag based on a comparison result of the text weight value and a preset text threshold, a comparison result of the image weight value and a preset image threshold, and a comparison result of the metadata weight value and a preset metadata threshold;

[0021] Based on the state combination of the rule network trigger flag, the semantic network trigger flag and the domain network trigger flag, a dynamic processing path sequence is generated in the order that the rule network has a higher priority than the semantic network, and the semantic network has a higher priority than the domain network. The execution result of each processing path in the dynamic processing path sequence is used as the input condition for the next processing path.

[0022] In a preferred embodiment, when the rule network trigger flag is valid, the archive text data is filtered layer by layer based on the risk keyword matching rules in the rule network, and the risk keyword matching rules include keyword whitelist, keyword blacklist and regular expression combination matching;

[0023] When the semantic network trigger flag is valid, the implicit semantic association between text and image is extracted through the cross-modal association model, and the cross-modal sensitive information transmission path is identified based on the association strength threshold;

[0024] When the domain network trigger flag is valid, a spatial overlap analysis is performed based on the geographic coordinate field in the archive metadata and the list of sensitive regions in the regional feature database to calculate the regional matching confidence.

[0025] In a preferred embodiment, S3 includes:

[0026] Analyze the new and abolished risk keywords in the regulatory update directives and generate a keyword update list;

[0027] Weight parameters of the risk keyword matching rules associated with the keyword update list positioning rule network, the weight parameters include priority coefficient and matching sensitivity;

[0028] Construct a weight parameter optimization model with gradient projection constraints. The constraints include the upper limit of the weight parameter variation and compliance verification rules.

[0029] The positioning weight parameters are iteratively optimized through the gradient projection algorithm. The iterative termination condition is that the parameter change is less than the convergence threshold or the maximum number of iterations is reached;

[0030] Update the optimized weight parameters to the rule network, perform validation set testing on the updated rule network, and activate the updated rule network after passing the test;

[0031] When the validation set test fails, the system rolls back to the previous valid weight parameters and triggers the manual intervention process, which includes manual calibration of weight parameters and analysis of abnormal logs.

[0032] In a preferred embodiment, S4 includes:

[0033] Extract the auditor's professional field labels, historical audit records, and performance scores to generate the initial eigenvectors of the multi-dimensional professional feature matrix;

[0034] Normalize the initial feature vectors to eliminate the dimensional differences of different feature dimensions and generate a normalized multi-dimensional professional feature matrix;

[0035] Encode the cross-modal sensitive association data identified by the semantic network into sensitive feature vectors;

[0036] Calculate the cosine similarity between the sensitive feature vector and each feature vector in the normalized multi-dimensional professional feature matrix to generate a similarity matching result;

[0037] Based on the similarity matching results and the preset qualification threshold, the file data is dynamically allocated to the review node of the corresponding qualification;

[0038] When the similarity matching result is lower than the qualification threshold, the cross-node collaborative review process is triggered. In the collaborative review process, the main review node and the auxiliary node jointly generate the review conclusion according to the weight ratio.

[0039] In a preferred embodiment, the sensitive feature vector includes text sensitive entities, image sensitive areas and association strength values.

[0040] In a preferred embodiment, S5 includes:

[0041] Collect audit process data assigned to audit nodes, including task processing time, manual confirmation node identification, and paper transmission path;

[0042] Identify manual confirmation links and paper transmission paths in the audit process as redundant nodes based on value stream analysis;

[0043] Build a lightweight audit process optimization model, replace manual confirmation links with automated rule verification nodes, and replace paper transmission paths with blockchain digital signature transmission channels;

[0044] Verify the integrity of the optimized process through process simulation. Verification conditions include meeting the task processing time compression ratio and passing the data hash consistency check.

[0045] The lightweight audit process that has passed verification will be released in a phased manner, and the original process will be replaced in stages according to the load weight of the audit node;

[0046] When a process exception occurs during a grayscale release, a rollback mechanism is triggered and an optimization defect report is generated.

[0047] In a preferred embodiment, the defect report includes the exception type, triggering conditions, and repair suggestions.

[0048] In a preferred embodiment, S6 includes:

[0049] Perform data segmentation processing on multimodal feature vectors, dynamically allocated processing paths, and lightweight audit process operation records to generate a structured evidence data set;

[0050] The structured evidence data set is split into multiple data blocks according to the preset block partitioning strategy, and each data block is attached with the hash value of the previous block to generate a chain association;

[0051] Data blocks are written to the distributed ledger in parallel through blockchain nodes, and smart contracts are called during writing to verify data integrity and permission compliance;

[0052] Generate a Merkle tree root hash value containing a write timestamp on the blockchain network and use the Merkle tree root hash value as the evidence chain identifier;

[0053] A secondary timestamp signature is performed on the evidence chain identifier based on the timestamp service to generate an unalterable timestamp evidence chain.

[0054] When the stored data needs to be verified, the data integrity and timestamp authenticity are verified by backtracing the Merkle tree path and matching the hash value.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. By integrating cross-modal semantic understanding and dynamic path allocation mechanism, the intelligent level of archive open identification has been significantly improved; the multimodal feature vector constructed based on the hybrid expert model can deeply analyze the implicit associations between text, images and metadata, and dynamically allocate processing paths in combination with attention weights, so that the rule network, semantic network and domain network form a collaborative verification closed loop; it not only breaks through the limitations of traditional single-modal processing, but also through the precise division of labor between cross-modal sensitive information identification and regional feature matching, while reducing the reliance on manual review, it realizes full-dimensional risk screening of complex archival data, effectively solving the core pain points of high misjudgment rate and low cross-departmental collaboration efficiency; the rule network optimization method with gradient projection constraints enables keyword matching rules to respond to regulatory changes in real time, avoiding the resource consumption of traditional full-model iteration, and ensuring the timeliness and compliance of the audit strategy.

[0057] 2. By introducing a multi-dimensional professional feature matrix of auditors, manual experience is precipitated into a quantifiable capability model. By dynamically matching it with sensitive data output by the semantic network, accurate adaptation of task allocation and personnel qualifications is achieved, maximizing the effectiveness of professional audits. Lightweight process reconstruction driven by value stream analysis eliminates redundant manual links and paper transmission nodes, and builds an efficient channel for automated rule verification and digital signature transmission, which greatly improves the response speed of cross-departmental collaboration. Finally, blockchain evidence storage technology solidifies the entire multimodal data processing process into an unalterable chain of evidence, and the timestamp signature and Merkle tree verification mechanism form a double anti-counterfeiting barrier, which not only improves the transparency of audits but also provides a complete and reliable technical closed loop for subsequent tracing. Through the overall deep integration of data analysis, rule iteration, human-computer collaboration and trusted evidence storage, the problem of balancing accuracy, real-time performance and compliance in the open archives identification scenario is systematically solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1The present invention is a flowchart of a method for intelligent open identification of archives based on a large model. DETAILED DESCRIPTION

[0059] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0060] Example: Figure 1 The present invention provides a method for intelligent open identification of archives based on a large model, comprising the following steps:

[0061] S1. Perform cross-modal semantic analysis on the text data, image data and metadata of the archives, and generate multimodal feature vectors based on the hybrid expert model;

[0062] S2. dynamically assigning the archival data to at least one processing path among a rule network for performing risk keyword matching, a semantic network for identifying cross-modal sensitive associations, and a domain network for matching regional features based on the attention weight distribution of each data modality in the multimodal feature vector;

[0063] S3. Based on the regulatory update directive, the weight parameter optimization method based on gradient projection constraints is used to locally adjust the weight parameters related to risk keyword matching in the rule network;

[0064] S4. Construct a multi-dimensional professional feature matrix for auditors, perform similarity matching between the cross-modal sensitive correlation data identified by the semantic network and the multi-dimensional professional feature matrix, and assign them to the audit nodes corresponding to the qualifications;

[0065] S5. Perform value stream analysis on the audit process assigned to the audit node, remove manual confirmation steps and paper transmission nodes, and generate a lightweight audit process;

[0066] S6. Write the multimodal feature vector, the dynamically allocated processing path, and the operation record of the lightweight audit process into the blockchain to generate a timestamped evidence chain.

[0067] S1. Perform cross-modal semantic analysis on the text data, image data, and metadata of the archive, and generate multimodal feature vectors based on the hybrid expert model. The specific implementation is as follows:

[0068] Cross-modal alignment is performed on the text data, image data, and metadata of the archives. Cross-modal alignment is achieved by calculating the semantic relevance between the text description and the image content. Specifically, a pre-trained cross-modal embedding model is used. The training data for the cross-modal embedding model comes from a public image-text alignment dataset. The training goal is to minimize the cosine distance between the text embedding vector and the corresponding image embedding vector. The dimensions of the text embedding vector and the image embedding vector are both 512. The semantic relevance is obtained by calculating the cosine similarity between the text embedding vector and the image embedding vector. The calculation formula is the dot product of the two vectors divided by the product of the vector moduli. The calculation result ranges from -1 to 1.

[0069] The semantic relevance threshold was determined through experiments on a validation set consisting of historical archival data, including manually annotated text and image association labels. The validation set accuracy target was set to no less than 95% of the historical manual review accuracy. A grid search was used to iteratively test the values ​​within the range of 0.5 to 0.8 with a step size of 0.05, and the final threshold was determined to be 0.65. When the semantic relevance falls below the threshold, a manual annotation process is triggered to supplement the association labels. This process notifies the reviewer by calling the review system interface to complete the annotation within 24 hours. The cross-modal alignment feature vector is generated by combining the text embedding vector, the image embedding vector, and the semantic relevance. This combination is achieved by concatenating the text embedding vector and the image embedding vector into a 1024-dimensional vector along the channel dimension. The semantic relevance is then normalized to a weight coefficient and weightedly fused with the concatenated vector. The weighting formula is: the normalized weight coefficient multiplied by the value of each dimension of the concatenated vector.

[0070] The cross-modal aligned feature vectors are input into the text expert network, image expert network, and metadata expert network in the hybrid expert model to extract deep semantic features for the text, spatial features for the image, and spatiotemporal features for the metadata, respectively. The text expert network adopts a bidirectional long short-term memory (LSTM) network architecture. The network input is the text embedding vector portion of the cross-modal aligned feature vector. The number of hidden layers in the LSTM network is dynamically adjusted based on the training data size. The initial number of hidden units is 256, and the activation function is the hyperbolic tangent function. The network output is a 512-dimensional deep semantic feature of the text, capturing the sensitive entities and contextual semantic relationships in the text. The image expert network adopts a convolutional neural network architecture. The network input is the image embedding vector portion of the cross-modal aligned feature vector. The convolutional neural network consists of five convolutional layers with a 3×3 kernel size, a stride of 1, and symmetric padding. The pooling layer uses a max pooling operation with a kernel size of 2×2. The network output is a 512-dimensional spatial feature of the image, identifying the spatial distribution and visual semantics of sensitive regions in the image.

[0071] The metadata expert network adopts a fully connected network structure. The network input is the semantic relevance weight value and archive metadata in the cross-modal alignment feature vector. The archive metadata includes the creation timestamp, geographic coordinates and storage path hash value. The input layer dimension of the fully connected network is the sum of the semantic relevance weight value dimension and the number of metadata fields. The hidden layer dimension is rounded down to half of the input layer dimension. The activation function is a linear rectification function. The network output is a 512-dimensional metadata spatiotemporal feature, which represents the spatiotemporal correlation between the archive creation time, geographic location and storage path.

[0072] Multimodal feature fusion is performed on text deep semantic features, image spatial features, and metadata spatiotemporal features. Weight coefficients for different modal features are calculated using a multi-head attention mechanism. The implementation of the multi-head attention mechanism involves linearly transforming the text deep semantic features, image spatial features, and metadata spatiotemporal features to generate a query vector, a key vector, and a value vector. The dimensions of the linear transformation matrix are input feature dimensions × output feature dimensions, and the output feature dimensions are determined by the number of attention heads. Each attention head independently calculates the dot product similarity between the query vector and the key vector. The dot product result is scaled by dividing it by the square root of the feature vector dimension to 64, and the scaling factor is the inverse of the feature vector dimension under the square root (0.125). The scaled result is normalized into weight coefficients using a Softmax function. The temperature parameter of the Softmax function is set to a fixed value of 1.0 to ensure the stability of the weight distribution. The input of the Softmax function is the scaled dot product result, and the output is the normalized weight coefficient of each modal feature.

[0073] Based on the weight coefficients, a weighted summation of the text deep semantic features, image spatial features, and metadata spatiotemporal features is performed to generate a multimodal feature vector. During the weighted summation process, the weight coefficient of each modal feature is multiplied element-by-element by the corresponding feature vector. The weighted feature vectors of the three modalities are concatenated according to the channel dimension to form the final multimodal feature vector. The dimension of the multimodal feature vector is determined by the sum of the dimensions of the feature vectors of each modality. The dimensions of the text deep semantic features, image spatial features, and metadata spatiotemporal features are all 512-dimensional. After concatenation, a 1536-dimensional multimodal feature vector is generated, which is used for dynamic allocation and processing path decision-making in subsequent steps. For archival data with missing metadata, the metadata expert network automatically fills in default spatiotemporal features. The default value for the creation timestamp is January 1, 1970, the default value for the geographic coordinates is a null value marker, and the default value for the storage path hash value is a zero vector.

[0074] S2. Based on the attention weight distribution of each data modality in the multimodal feature vector, dynamically assign the archival data to at least one processing path among a rule network for performing risk keyword matching, a semantic network for identifying cross-modal sensitive associations, and a domain network for matching regional features. The specific implementation is as follows:

[0075] The attention weight distributions of text, image, and metadata in the multimodal feature vector are normalized to generate text weight values, image weight values, and metadata weight values. This normalization uses a linear normalization method. The specific formula is as follows: the original attention weight value of the text modality is divided by the sum of the original attention weight values ​​of the three modalities to obtain the text weight value; the original attention weight value of the image modality is divided by the sum of the original attention weight values ​​of the three modalities to obtain the image weight value; and the original attention weight value of the metadata modality is divided by the sum of the original attention weight values ​​of the three modalities to obtain the metadata weight value.

[0076] The initial values ​​of text weight value, image weight value and metadata weight value are determined by the output of the multi-head attention mechanism of the hybrid expert model. The training process of the hybrid expert model is based on the historical archive dataset. The historical archive dataset contains manually annotated text importance labels, image key area labels and metadata relevance labels. The dataset is divided into 80% training set, 10% validation set and 10% test set.

[0077] The preset text threshold, preset image threshold and preset metadata threshold are determined through validation set experiments. The validation set consists of 1,000 independent archival data, and the data distribution is consistent with the training set. The accuracy, recall rate and F1 score of different threshold combinations are tested in the range of 0.3 to 0.7 with a step size of 0.05 through grid search. Finally, the threshold combination with the highest comprehensive F1 score is selected, and the text threshold is determined to be 0.55, the image threshold is 0.45, and the metadata threshold is 0.4.

[0078] When the text weight is greater than or equal to the text threshold, a rule network trigger flag is generated. When the image weight is greater than or equal to the image threshold, a semantic network trigger flag is generated. When the metadata weight is greater than or equal to the metadata threshold, a domain network trigger flag is generated. Archival data with a weight below the threshold is marked as a low-priority processing object and stored in a buffer queue. The maximum capacity of the buffer queue is set to 1000. After the capacity is exceeded, old data is eliminated according to the first-in, first-out policy.

[0079] When the rule network trigger flag is valid, the archive text data is filtered layer by layer based on the risk keyword matching rules in the rule network. Risk keyword matching rules include keyword whitelists, keyword blacklists, and regular expression combination matching. The keyword whitelist consists of high-frequency non-sensitive words from historical security archives. The screening criteria for high-frequency non-sensitive words are words that rank in the top 10% of frequency in the training set and have not been manually labeled as sensitive. The update frequency is once a week, and the integrity of the vocabulary is ensured through hash verification during updates. The keyword blacklist consists of sensitive words that are expressly prohibited by the "Data Security Law of the People's Republic of China" and the "Personal Information Protection Law of the People's Republic of China". The inclusion of sensitive words is based on the "Information Security Technology Personal Information Security Specification" issued by the National Standardization Administration. The update frequency is real-time synchronization with regulatory revisions, and incremental synchronization is achieved through version number comparison during updates.

[0080] Regular expression combination matching includes ID card number matching rules (such as "\d{17}[\dXx]"), phone number matching rules (such as "1[3-9]\d{9}"), and timestamp matching rules (such as "\d{4}-\d{2}-\d{2}"). The priority of regular rules decreases in the order of ID card number, phone number, and timestamp. The layer-by-layer filtering process sequentially performs whitelist safe word matching, blacklist sensitive word exclusion, and regular expression pattern screening. Whitelist matching is implemented through a full-word matching algorithm, which uses a dual-array Trie tree structure to accelerate queries; blacklist exclusion is implemented through inverted index retrieval, and the inverted index is constructed once a day; regular expression screening is implemented through a finite state automaton, and the construction of the finite state automaton is based on a deterministic finite automaton optimization algorithm.

[0081] If the text data passes the whitelist match and does not trigger the blacklist exclusion, regular expression screening will continue; if an alarm is triggered at any stage, subsequent screening will be terminated and the file will be marked as a high-risk file. The marking information of the high-risk file includes the trigger rule type, matching location and risk level. The risk level is divided into three levels, from high to low, corresponding to immediate interception, manual review and additional marking.

[0082] When the semantic network trigger flag is valid, the cross-modal association model extracts implicit semantic associations between text and images and identifies cross-modal sensitive information transmission paths based on an association strength threshold. The cross-modal association model is pre-trained on a publicly available image-text alignment dataset. The training data consists of 120,000 image-text pairs from the MS-COCO dataset. The training objective is to minimize the cosine distance of positive pairs and maximize the cosine distance of negative pairs. The optimization algorithm uses AdamW, with an initial learning rate of 3e-5, a batch size of 32, and 50 epochs. The model outputs an association strength value between text and image. The original association strength value ranges from -1 to 1 and is mapped to the range from 0 to 1 using the linear transformation formula Snorm = (Sraw + 1) / 2. Sraw represents the raw association strength value output by the cross-modal association model, ranging from -1 to 1; Snorm represents the normalized association strength value after the linear transformation, ranging from 0 to 1.

[0083] The association strength threshold was determined using the receiver operating characteristic (ROC) curve of the validation set, which contained 500 manually annotated cross-modal sensitive association profiles. A threshold of 0.65 was selected as the decision threshold, maximizing the ratio of the true positive rate to the false positive rate. The implicit semantic association extraction process involved encoding the text description and image content into 512-dimensional feature vectors, respectively. The text encoder used the BERT-base model, while the image encoder used the ResNet-50 model. The cosine similarity between the two was calculated as the association strength value. Cosine similarity is calculated as the dot product of the two vectors divided by the product of their moduli. The vector moduli were calculated using the Euclidean norm.

[0084] When the correlation strength value is greater than or equal to the threshold, a cross-modal sensitive information transmission path is determined to exist. For example, if the text description "military facilities" is highly correlated with the outline of a closed area in the image, an alarm is triggered. When the correlation strength value is lower than the threshold, it is marked as no sensitive association. For cases where the correlation strength value is within the threshold ±0.1, a secondary verification process is automatically triggered. The secondary verification process calls the manual review interface and returns the review results within 24 hours.

[0085] When the domain network trigger flag is valid, a spatial overlap analysis is performed based on the geographic coordinate fields in the archive metadata and the list of sensitive areas in the regional feature database to calculate the regional match confidence. The list of sensitive areas includes the controlled geofences specified in the National Security Law. Each geofence is composed of a polygon defined by at least four longitude and latitude coordinates, with an accuracy of six decimal places and a coordinate system of standard WGS84.

[0086] Spatial overlap analysis uses the ray method to determine whether the archive's geographic coordinates are located within the fenced polygon. This method involves emitting rays from the point to be measured in any direction and counting the number of intersections between the rays and the polygon boundary. If the number of intersections is odd, the point is considered inside the polygon; if it is even, the point is considered outside. The closest distance between the archive's coordinates and the fence boundary is calculated using the spherical distance formula, which is expressed in kilometers. The formula is the Haversine formula, which takes the longitude and latitude coordinates of two points as input and outputs the great-circle distance between the two points.

[0087] The calculation formula for regional matching confidence is the ratio of the area of ​​the overlapping area to the total area of ​​the sensitive area fence. The area calculation adopts the spherical geometry algorithm. The implementation of the spherical geometry algorithm is based on the open source library GeographicLib, and the calculation result is rounded to two decimal places. When the matching confidence is greater than or equal to the preset confidence threshold of 0.75, the archive is determined to be a sensitive area-related archive; when the matching confidence is lower than the threshold, it is marked as having no regional risk. For archives with missing geographic coordinates in the metadata, the default coordinate value (longitude 0.0, latitude 0.0) is automatically filled in and the manual review process is triggered. The response time for manual review does not exceed 24 hours, and the manual review results are fed back to the review system through the message queue.

[0088] Based on the combined states of the rule network trigger flag, the semantic network trigger flag, and the domain network trigger flag, a dynamic processing path sequence is generated, with the rule network taking precedence over the semantic network, which in turn takes precedence over the domain network. The dynamic processing path sequence generation logic is as follows: If the rule network trigger flag is valid, the rule network processing path is executed first. High-risk files in the rule network processing results are directly placed in the manual review queue, with a maximum concurrent processing capacity of 100 files per second. Low-risk files continue to execute the semantic network processing path. If the rule network is not triggered but the semantic network trigger flag is valid, the semantic network processing path is executed. High-risk files in the semantic network processing results are placed in the manual review queue, while low-risk files continue to execute the domain network processing path. If only the domain network trigger flag is valid, the domain network processing path is executed independently, and the processing results are directly written into the audit evidence chain.

[0089] The output data format for each processing path is a JSON structure, containing the profile ID, risk level, triggering rule details, and processing timestamp. The processing timestamp is generated with millisecond accuracy and is standardized to UTC. The timestamp server uses the NTP protocol to synchronize with the National Time Service Center's time source. For system timeouts or data anomalies that occur during the execution of a processing path, an exception handling mechanism is activated. This includes automatic retries (maximum retries: three), downgrade processing (skipping the current processing path), and error logging. The error log contains fields that include the exception type, occurrence time, and associated profile ID.

[0090] Step S2 solves the problems of missed detection of cross-modal sensitive information and regional matching deviation caused by single threshold decision-making in traditional audit methods through a multi-threshold dynamic triggering mechanism and hierarchical processing path design. Compared with the existing technology that adopts fixed rule routing, this step independently triggers different networks based on normalized weight values ​​to achieve refined classification and processing of multimodal archival data: the rule network prioritizes filtering explicit risks, the semantic network captures cross-modal implicit associations, and the domain network accurately matches regional characteristics. The three form a collaborative verification chain through priority sorting, which not only avoids the update lag caused by over-reliance on manual rules, but also overcomes the semantic fragmentation defect of single model processing multimodal data. Through the weight threshold grading and dynamic combination of paths, the coverage and accuracy of sensitive information identification are improved while ensuring audit efficiency, achieving dual optimization of regulatory compliance and business continuity.

[0091] S3. Based on the regulatory update instructions, the weight parameter optimization method based on gradient projection constraints is used to locally adjust the weight parameters related to risk keyword matching in the rule network. The specific implementation is as follows:

[0092] Parse the new and repealed risk keywords in the regulatory update directive to generate a keyword update list. The input format for the regulatory update directive is structured text. The parsing rules for structured text are based on predefined regular expression templates. The regular expression templates contain matching patterns for new keywords and repealed keywords. The matching pattern for new keywords is defined as "[New](.+?)[ / New]", and the matching pattern for repealed keywords is defined as "[Repealed](.+?)[ / Repealed]".

[0093] During the parsing process, the regular expression engine is used to extract the list of new keywords and the list of abolished keywords to generate a keyword update list. The data structure of the keyword update list is a set of key-value pairs, where the key is the keyword hash value, the hash algorithm uses the MD5 algorithm, and the value is the keyword type tag (new / abolished) and the effective timestamp. For keywords with format errors or semantic conflicts that appear during the parsing process, a manual review process is triggered. The manual review process is assigned to designated reviewers through the review system work order interface. The review response time threshold is 2 hours. If it is not processed within the timeout, it will be automatically escalated to the system administrator.

[0094] The weight parameters for matching rules associated with risk keywords in the keyword update list positioning rule network are: priority coefficient and matching sensitivity. The priority coefficient controls the matching order of different risk keywords. The initial value of the priority coefficient is dynamically set based on the keyword's frequency of occurrence in historical audit data; higher frequency indicates a higher priority coefficient. Matching sensitivity adjusts the strictness of keyword matching. The initial value of matching sensitivity is set based on the keyword's semantic risk level; higher risk levels indicate higher matching sensitivity.

[0095] The positioning process is achieved by traversing the parameter index table of the rule network. The key field of the parameter index table is the risk keyword hash value. The hash value generation rule is consistent with the keyword update list. For new keywords, a new entry is created in the parameter index table and the priority coefficient is initialized to the default value of 0.5 and the matching sensitivity is initialized to the default value of 0.7. The default values ​​are set based on the median of historical parameter statistics. For discontinued keywords, the priority coefficient and matching sensitivity of the corresponding entry are marked as invalid. The weight parameters of the invalid state are not included in the subsequent optimization process. The invalid state is marked by setting a disable flag in the parameter index table.

[0096] Construct a weight parameter optimization model with gradient projection constraints. The constraints include an upper limit on the range of weight parameter changes and compliance verification rules. The upper limit of the weight parameter change range is determined through statistical analysis of historical parameter adjustment records. The statistical method is to calculate the standard deviation of the absolute change of the last 10 parameter adjustments. The upper limit is set to the mean plus twice the standard deviation to prevent system instability caused by sudden parameter changes during the optimization process. Compliance verification rules include that new keywords must pass whitelist verification and that revoked keywords must exist in the current parameter index table. Whitelist verification is achieved by querying a predefined security vocabulary, and the update frequency of the security vocabulary is synchronized with the regulatory release cycle.

[0097] The mathematical model of the gradient projection constraint is to transform the parameter optimization problem into a constrained convex optimization problem. The projection operation ensures that the parameter value after each iteration meets the constraint conditions. The projection operation is implemented by truncating the parameter value that exceeds the constraint range to the nearest boundary value.

[0098] The positioning weight parameters are iteratively optimized through the gradient projection algorithm. The iterative termination condition is that the parameter change is less than the convergence threshold or the maximum number of iterations is reached. The implementation process of the gradient projection algorithm includes: calculating the gradient of the loss function of the weight parameter. The loss function is defined as the weighted sum of the matching omission rate of the newly added keywords and the mismatch rate of the abolished keywords. The weight coefficient is dynamically adjusted according to the regulatory priority. The regulatory priority is determined by the urgency of the effective timestamp; projecting the gradient vector into the compliance constraint space to generate the compliance gradient direction. The mathematical representation of the compliance constraint space is the feasible domain of linear inequality constraints; updating the weight parameter along the compliance gradient direction. The update step size is determined by linear search. The initial step size is set to 0.1 and the maximum step size does not exceed 0.5. The step size adjustment rule is that if the current step size causes the loss function to increase, the step size is halved and recalculated.

[0099] The convergence threshold is set experimentally as the L2 norm of the parameter change is less than 1e-4. The calculation formula of the L2 norm is the square root of the sum of the squares of the parameter changes; the maximum number of iterations is set to 100. After the number of iterations is exceeded, the optimization is forced to terminate and marked as failed. The optimization failure flag is written to the system log and an alarm notification is triggered.

[0100] The optimized weight parameters are updated to the rule network, and the updated rule network is tested on the validation set. Once the test passes, the updated rule network is activated. The validation set consists of 1,000 historical archival data items, including manually annotated true risk labels. The validation set is divided into 30% high-risk files, 50% medium-risk files, and 20% low-risk files. Validation set test metrics include precision, recall, and F1 score. Precision is calculated as the number of correctly identified high-risk files divided by the total number of test files. Recall is calculated as the number of correctly identified high-risk files divided by the total number of true high-risk files. The F1 score is calculated as the harmonic mean of precision and recall.

[0101] The test passing standard is that the F1 score is not less than 95% of the previous parameters. If the test passes, the updated rule network is activated; the activation process adopts a hot update mechanism, which is implemented through double buffering technology. The specific process of double buffering technology is: create a new parameter buffer and an old parameter buffer in the memory, and gradually migrate the new parameters from the new buffer to the old buffer during the update. During the migration, the service request is responded to by the old buffer, and the buffer pointer is switched after the migration is completed.

[0102] If the validation set test fails, the system rolls back to the previously valid weight parameters and triggers manual intervention, which includes manual calibration of the weight parameters and analysis of abnormality logs. The rollback operation is implemented through a version control system, which records the hash value and timestamp of each parameter update. During a rollback, the system retrieves the previously valid version and fully replaces the current parameters. The storage backend of the version control system is a distributed database with three data replicas to ensure disaster recovery.

[0103] In the manual intervention process, manual calibration of weight parameters is completed through a visual configuration interface. The visual configuration interface displays the comparison data before and after parameter optimization and the verification set test results. The comparison data includes a priority coefficient change trend chart and a matching sensitivity distribution histogram. Abnormal log analysis includes parsing the reasons for optimization failure (such as gradient explosion, constraint conflict, etc.) and generating a repair suggestion report. The repair suggestion report is automatically generated through natural language processing technology. The natural language processing model is pre-trained based on the Transformer architecture. The output format is structured text and is pushed to the auditor's workbench. For scenarios where optimization fails three times in a row, the rule network architecture reconstruction process is automatically triggered. The reconstruction process includes rebuilding the parameter index table and resetting the constraint conditions. The steps for rebuilding the parameter index table are to clear the existing entries and reload the latest regulatory keywords. The steps for resetting the constraint conditions are to recalculate the upper limit of the weight parameter change range and the compliance verification rules.

[0104] S4. Construct a multi-dimensional professional feature matrix for auditors, perform similarity matching between the cross-modal sensitive correlation data identified by the semantic network and the multi-dimensional professional feature matrix, and assign them to the audit nodes of the corresponding qualifications. The specific implementation is as follows:

[0105] The auditor's professional domain labels, historical audit records, and performance scores are extracted to generate the initial eigenvectors of the multidimensional professional feature matrix. Professional domain labels are determined by the system's pre-set domain classification system, which includes three categories: file classification assessment, sensitive information identification, and regional compliance review. The weight of each label is set based on the auditor's qualification certificate level. Qualification certificate levels are categorized as junior, intermediate, and senior, with corresponding weight coefficients of 0.6, 0.8, and 1.0, respectively.

[0106] Historical audit records include the number of files processed within the past year, their accuracy rate, and average processing time. The data source is the audit system database. Data cleaning rules eliminate abnormal records with audit times exceeding 24 hours or accuracy rates below 60%. The cleaned data is statistically analyzed monthly to generate standardized indicators. The performance score is calculated by weighting the audit quality score and efficiency score, with the quality score accounting for 70% and the efficiency score accounting for 30%. The weighting coefficient is determined through annual management reviews. The initial feature vector is generated by concatenating the professional field label weights, historical audit record indicators, and performance scores by dimension, forming a vector containing 10 feature dimensions. The vector dimension order is fixed as follows: professional field weight (3 dimensions), number of files (1 dimension), accuracy rate (1 dimension), average processing time (1 dimension), quality score (1 dimension), efficiency score (1 dimension), total performance score (1 dimension), and backup dimension (1 dimension).

[0107] The initial eigenvectors are standardized to eliminate the dimensional differences between different feature dimensions and generate a normalized multi-dimensional professional feature matrix. The Z-score standardization method is used for standardization. The specific formula is to subtract the mean of each feature dimension from the feature value and divide it by the standard deviation. The mean and standard deviation are calculated based on the feature data of all auditors, and the calculation cycle is once a month. The way to handle missing values ​​is as follows: if a feature dimension is missing, it is filled with the mean of all personnel in that dimension; if more than three feature dimensions are missing, it is marked as invalid data and the manual completion process is triggered. The manual completion process is assigned to the designated administrator through the audit system work order interface, and the completion response time threshold is 4 hours. The normalized multi-dimensional professional feature matrix is ​​stored in a distributed database. The database adopts a columnar storage structure and enables data compression algorithm. The compression algorithm is Zstandard, and the compression level is set to 3 to balance storage efficiency and query performance.

[0108] The cross-modal sensitive association data identified by the semantic network is encoded into a sensitive feature vector, which contains text-sensitive entities, image-sensitive regions, and association strength values. Text-sensitive entities are extracted using a named entity recognition model that utilizes a pre-trained BERT-base architecture. Entity types include names of people, places, organizations, and sensitive keywords. The extracted entity list is encoded into a 256-dimensional vector using a word embedding model. The word embedding model is trained on a public general corpus and fine-tuned using historical archival data to adapt to domain semantics.

[0109] Sensitive regions of the image are extracted using a target detection model based on the YOLOv5 architecture. Detection categories include faces, license plates, geographic landmarks, and sensitive objects. The detection results are generated into 256-dimensional vectors using a regional feature encoder based on the ResNet-50 model. The output layer is replaced with a global average pooling layer to adapt to the feature dimension. The association strength value is output by the cross-modal association model of the semantic network. The original range of the association strength value is -1 to 1, and it is mapped to the range of 0 to 1 using a linear transformation formula: the association strength value plus 1 divided by 2. The final dimension of the sensitive feature vector is 513 (256+256+1), stored in a single-precision floating-point array format, with 64-byte array memory alignment to optimize vector computation performance.

[0110] The cosine similarity between the sensitive feature vector and each eigenvector in the normalized multidimensional professional feature matrix is ​​calculated to generate a similarity match result. The cosine similarity calculation process involves performing a dot product between the sensitive feature vector and each eigenvector in the multidimensional professional feature matrix. The dot product result is then divided by the product of the two vectors' moduli, calculated using the Euclidean norm. The similarity match result is output as a sorted list, sorted from highest to lowest similarity value. Each entry contains the reviewer ID, similarity value, and the contribution of the matching feature dimension. The contribution of the matching feature dimension is determined by calculating the contribution ratio of each dimension's eigenvalue to the similarity result. The contribution ratio formula is the percentage of the product of each dimension's features to the total dot product value. Entries with the same similarity value are sorted in descending order by reviewer performance score. If the performance score is the same, they are sorted in ascending order by the average processing time of the historical review records. The maximum number of entries displayed in the sorted list is set to 50. Any entries exceeding this limit are loaded in pages to reduce front-end rendering pressure.

[0111] Based on similarity matching results and preset qualification thresholds, archive data is dynamically assigned to audit nodes with corresponding qualifications. The qualification threshold is dynamically adjusted based on the audit node's professional competence level, which is divided into three levels: the qualification threshold for first-level audit nodes is set at the top 10% in similarity ranking, the top 30% for second-level audit nodes, and the top 50% for third-level audit nodes. The threshold adjustment rule is as follows: When the sensitivity level of the audited archive is high, the qualification threshold automatically increases by 10%. Sensitivity is determined by the association strength value output by the semantic network, with association strength values ​​greater than 0.8 being marked as high sensitivity. When the system is under high load (with more than 1,000 concurrent tasks), the qualification threshold decreases by 5% to balance the load. The allocation logic is as follows: the first audit node from the similarity-ranked list that meets the qualification threshold is selected as the primary audit node. If no node in the list meets the threshold, the cross-node collaborative review process is triggered. During the allocation process, the task queue length of the audit node is updated in real time. Nodes with a task queue length exceeding 20 items are automatically marked as busy and temporarily removed from the allocation list.

[0112] When the similarity match result falls below the qualification threshold, a cross-node collaborative review process is triggered. During this collaborative review process, the lead review node and auxiliary nodes jointly generate a review conclusion based on a weighted ratio. The lead review node is the node with the highest score in the similarity ranking list. Up to two auxiliary nodes are selected from the highest similarity score, with an auxiliary node weight of 50% of the lead review node. The collaborative review decision-making rule is a weighted voting mechanism, with the lead review node having a weight of 1 and the auxiliary nodes having a weight of 0.5. The final review conclusion is determined when the weighted score exceeds the preset decision threshold of 0.7. If the weighted score does not reach the threshold, the case is escalated to an expert committee for review. The expert committee is composed of three senior reviewers and uses a majority vote to reach the final review conclusion. The majority vote requirement for approval is unanimous agreement between at least two experts. During the collaborative review process, all node operation records and decision-making basis are synchronized in real time to the blockchain evidence node. The evidence data includes the reviewer ID, timestamp, decision-making basis summary, and digital signature. The digital signature algorithm uses the ECDSA secp256k1 curve to ensure immutability. In the scenario where collaborative review times out (is not completed for more than 24 hours), the task reallocation process is automatically triggered, and nodes that have been idle in the last 30 minutes are given priority during reallocation.

[0113] S5. Perform value stream analysis on the audit process assigned to the audit node, remove manual confirmation links and paper transmission nodes, and generate a lightweight audit process. The specific implementation is as follows:

[0114] Audit process data assigned to audit nodes is collected, including task processing time, manual confirmation node identifiers, and paper transmission paths. Task processing time is extracted from the audit system log, which records the task start and end time for each audit node with millisecond-level timestamp accuracy and a unified UTC time zone. Manual confirmation node identifiers are obtained from the process engine's node metadata, which includes a manual operation tag and operator ID. The manual operation tag is determined by whether the node type field contains "manual review" or "supervisor approval." The paper transmission path is generated by digitally scanning paper document flow records. During scanning, optical character recognition (OCR) technology is used to identify document numbers and transfer signatory information. This OCR technology is implemented using a pre-trained Tesseract engine, and the recognition results are verified for format compliance using regular expressions. The collected process data is stored in a time series database, indexed by task ID to support fast retrieval. Data cleaning rules exclude anomalous records with processing times less than 1 second or exceeding 24 hours. Anomalous records are determined by the absolute value of the task processing time Z-score being greater than 3.

[0115] Based on value stream analysis, manual confirmation steps and paper transmission paths in the audit process are identified as redundant nodes. The specific method for value stream analysis is to construct a value stream map of the audit process, with nodes representing audit steps and edges representing task transfer paths. The value contribution of each node is calculated as the product of the inverse of the node's processing time and the task's criticality weight. The task's criticality weight is determined by the classification level of the document: ordinary, secret, and confidential, with corresponding weight coefficients of 1.0, 1.5, and 2.0, respectively. Manual confirmation steps are identified when the node metadata contains a manual operation marker and the value contribution is below a preset threshold, determined through historical data analysis and calculated by taking the 25th percentile of the value contribution of all manual nodes. Paper transmission paths are identified when the path contains a paper document scan record and the average transmission time exceeds three times the electronic transmission time. The benchmark for electronic transmission time is the 90th percentile of the historical average for similar tasks. The judgment results of redundant nodes are written into the process optimization task queue. The priority of the task queue is sorted according to the influence of the redundant nodes. The influence is determined by the weighted sum of value contribution and frequency of occurrence, and the weight coefficients are set to 0.6 and 0.4.

[0116] A lightweight audit process optimization model was constructed, replacing manual confirmation with automated rule-based verification nodes and replacing paper-based transmission paths with blockchain digital signature transmission channels. The automated rule-based verification nodes are implemented by loading risk keyword matching rules associated with the current audit task from the rule network. These matching rules include keyword whitelists, blacklists, and regular expression combinations. If a rule match is successful, the task is automatically marked as low-risk. Otherwise, a semantic network secondary verification is triggered, and the output of the semantic network secondary verification serves as the final judgment basis. The blockchain digital signature transmission channel is implemented by generating a hash value of the file content during task transmission using the SHA-256 hash algorithm. The generated hash value is then bound to the task metadata. The hash value and the audit node's digital signature are written to the blockchain evidence storage node using the ECDSA secp256k1 curve for digital signatures, with the private key stored in a hardware security module (HSM). The receiving node verifies the hash value consistency and signature validity, confirming the completion of the transmission. Transmission tasks that fail verification automatically trigger a retransmission mechanism, with a maximum of three retransmissions. The output of the optimization model is an updated audit process configuration table, which contains node types, associated rule bases, and transmission channel parameters. The configuration table version number is managed in the format of "major version.minor version.revision number".

[0117] The integrity of the optimized process is verified through process simulation. Verification conditions include the task processing time compression ratio meeting the requirements and passing the data hash consistency check. The specific steps of the process simulation are as follows: deploying the optimized audit process configuration table in a sandbox environment. The sandbox environment is isolated from the production environment network and has a resource quota of 4 CPU cores and 16GB of memory; importing a historical task data set for playback testing. The data set size is 1,000, and the data distribution is consistent with the production environment; the task processing time compression ratio is calculated as (original average duration - optimized average duration) / original average duration. The compression ratio threshold is set to no less than 20%, which is determined based on the minimum acceptable efficiency improvement in historical optimization cases; the data hash consistency check is achieved by comparing the file content hash values ​​before and after the transfer. The fault tolerance threshold of the consistency check is zero difference, and the difference detection mechanism uses a byte-by-byte comparison algorithm. The simulation test results generate a verification report, which includes a list of passed nodes, details of failed nodes, and performance improvement indicators. The performance improvement indicators include throughput improvement rate and error rate reduction rate. The throughput improvement rate is calculated as (optimized throughput - original throughput) / original throughput, and the error rate reduction rate is calculated as (original error rate - optimized error rate) / original error rate.

[0118] The lightweight audit process that has passed verification will be released in phases, replacing the original process in stages based on the audit node load weight. The phased release strategy is as follows: the release ratio is calculated based on the real-time load weight of the audit node. The load weight is dynamically calculated based on the node's CPU usage, memory utilization, and task queue length. The calculation formula is (CPU usage × 0.4 + memory utilization × 0.3 + task queue length × 0.3), and the result is normalized to the range of 0-1. The initial release ratio is set at 10%, and the scope is gradually increased by 20% daily until all nodes are switched over. During the release process, node processing performance and error rates are monitored in real time. Performance monitoring metrics include average response time and task timeout rate. The average response time threshold is set to no more than 120% of the historical average, and the task timeout rate threshold is set to no more than 5%. Error rate monitoring metrics include rule matching failure rate and signature verification failure rate, with failure rate thresholds set to no more than 2%. Monitoring data is collected via Prometheus and displayed in real time using the Grafana dashboard.

[0119] When a process anomaly occurs during a grayscale release, a rollback mechanism is triggered and an optimization defect report is generated. The defect report includes the anomaly type, triggering conditions, and remediation recommendations. The rollback mechanism is triggered if the node error rate exceeds the threshold for three consecutive samplings, or if the average processing time of a critical task exceeds 150% of the original process. Rollbacks are implemented through the version control system, which records the configuration table hash value and timestamp of each release. The target version to roll back is the process configuration table of the previous stable version. Version switching takes less than 5 minutes, and new task acceptance is temporarily frozen during the switch. The optimization defect report generation logic parses system logs for anomaly events, extracting the anomaly type (such as rule match failure, signature verification timeout) and triggering conditions (such as high concurrent load, network latency). Remediation recommendations are generated based on matching against a historical optimization case library, which is updated weekly. The matching algorithm uses cosine similarity search, and the results are a fusion of the remediation measures for the three most similar cases. Defect reports are assigned to the development team through the review system ticket interface. The remediation response time threshold is 48 hours. If unresolved after the timeout, they are automatically escalated to the technical lead.

[0120] S6. Write the multimodal feature vector, dynamically allocated processing path, and lightweight audit process operation records into the blockchain to generate a timestamped evidence chain. The specific implementation is as follows:

[0121] The multimodal feature vectors, dynamically assigned processing paths, and lightweight audit process operation records are segmented to generate a structured evidence dataset. The data segmentation rules are as follows: the multimodal feature vectors are segmented into 512-dimensional units, with each unit assigned a corresponding processing path identifier and an audit process operation timestamp. The processing path identifier is generated by concatenating the task ID and the processing path hash value. Dynamically assigned processing path data is grouped by task ID, with each group containing a sequence of path nodes and a processing result status code. The status code encoding rule is 0 for success, 1 for warning, and 2 for failure. The lightweight audit process operation records are aggregated by process instance ID. Each instance contains the operation type, executor ID, and an operation timestamp. The timestamp is generated with millisecond accuracy. The structured evidence dataset is formatted in JSON-LD, using the W3C standard data model definition to ensure semantic interoperability. During the data segmentation process, data exceeding the single block capacity limit is compressed using the Zstandard compression algorithm with a compression level of 5 to balance compression ratio and processing speed. The maximum size of the compressed data block is 1MB.

[0122] The structured evidence data set is split into multiple data blocks according to a preset block partitioning strategy. Each data block is appended with the hash value of the previous block to generate a chain association. The block partitioning strategy is as follows: the original data size of each data block does not exceed 1MB. If it exceeds this limit, it is automatically split into sub-blocks and a parent-child block index relationship is established. The parent-child index is implemented through a doubly linked list structure. The hash value of the previous block is generated by performing a SHA-256 hash operation on the entire data content of the previous block. The output length of the hash value is a 256-bit binary string, which is converted to hexadecimal encoding and stored. The chain association is implemented by writing the hash value of the previous block, the current block index number, and the block check code into the metadata header of the new block. The block check code is generated using the CRC32 algorithm and is generated once every 10 data packets. Data packets that fail the check trigger a retransmission process, with a maximum retransmission limit of 3 times.

[0123] Blockchain nodes write data blocks to the distributed ledger in parallel, invoking smart contracts to verify data integrity and permission compliance during writes. The blockchain node deployment architecture is a hybrid network consisting of both public cloud and privately deployed nodes. The inter-node consensus algorithm uses PBFT (Practical Byzantine Fault Tolerance), with a fault tolerance threshold of one-third of the total number of nodes. The smart contract's verification logic includes: verifying that the hash value of the data block matches the chained header; verifying that the digital signature of the operation record is generated by an authorized reviewer, whose public key certificate is pre-stored on the blockchain node; and verifying that the integrity hash of the multimodal feature vector is consistent with the original data recorded in the processing path. Blocks that fail verification are marked as invalid and placed in a retry queue with a maximum retry count of three. The retry interval increases exponentially, starting at 1 second and doubling with each retry.

[0124] A Merkle tree root hash value containing a write timestamp is generated on the blockchain network and used as the evidence chain identifier. The Merkle tree is constructed by using the hash values ​​of data blocks written in the same batch as leaf nodes, and then calculating the parent node hash values ​​layer by layer, pairwise, until a root hash value is generated. Timestamps are written by invoking a trusted timestamp service (TSA) when the root hash value is generated. The timestamp service's timing source is the National Time Service Center's NTP server, with millisecond-level timestamp accuracy. The timestamp data packet contains the root hash value, the TSA signature, and the timing source certificate. The format of the evidence chain identifier is "root hash value_timestamp_blockchain height." The blockchain height is determined by the total number of blocks currently on the chain. The height counter is stored as a 64-bit unsigned integer, and height updates are atomic operations to prevent concurrency conflicts.

[0125] The timestamp service performs a secondary timestamp signature on the evidence chain identifier, generating an unalterable, timestamped evidence chain. This secondary timestamp signature is implemented by binding the evidence chain identifier to the TSA's public key certificate to generate a digital signature. The signature algorithm uses the ECDSA secp256k1 curve, and the private key is hosted by a hardware security module (HSM). Access to the private key requires multi-factor authentication. The signature result and the original evidence chain identifier are encapsulated into an ASN.1-formatted data packet with a version of 1.0. The timestamped evidence chain is stored in the blockchain network's distributed file system, implemented using the IPFS protocol. The storage path is determined by a content-addressable hash value generated using the same hash algorithm as the block hash. Access to the evidence chain is controlled by a smart contract, limiting access to the complete data to authorized review nodes. The publicly verifiable portion contains only the evidence chain identifier and timestamp information. The query interface for public data is a RESTful API, with responses in JSON format.

[0126] When stored evidence data requires verification, data integrity and timestamp authenticity are verified through Merkle tree path backtracking and hash value matching. The verification process includes extracting the Merkle tree root hash value from the evidence chain identifier, recursively traversing the Merkle tree from parent nodes to leaf nodes, and verifying that the hash values ​​of child nodes are consistent with the records in the chain-linked header. Timestamp authenticity is verified by submitting the evidence chain identifier and signature data packet to the TSA service. The TSA returns the timing source certificate and signature validity status at the time of signing. The validity status is determined based on certificate chain verification and signature algorithm compliance checks. The verification results generate an audit report, which includes a list of verified blocks, details of abnormal data, and a timestamp validity mark. The abnormal data details include the index number of the block with a hash value mismatch and the corresponding original data snapshot. The snapshot data is retrieved from the IPFS storage by content hash value. The digital signature of the audit report is generated by the verification node, and the signature private key is stored in the HSM to ensure the credibility of the report.

[0127] Through the collaborative mechanism of cross-modal semantic analysis and dynamic path allocation, the technical limitations of single-modal processing and fixed rule routing in traditional archival auditing are overcome. Existing technologies typically rely on manual rules to define the processing flow of multimodal data, making it difficult to adaptively resolve implicit associations between text, images, and metadata. Steps S1 and S2 propose a dynamic allocation strategy based on a hybrid expert model and attention weights, dynamically routing multimodal feature vectors to rule, semantic, and domain networks based on semantic importance, achieving a precise division of labor for fine-grained sensitive information identification and regional compliance assessment. Step S3 introduces local optimization of the rule network with gradient projection constraints to address the rule update lag caused by traditional full-parameter retraining. Through regulatory directive-driven parameter fine-tuning, the rule base is adapted in real time while ensuring audit continuity. Steps S4 and S5 construct a closed-loop optimization of the auditor's competency matrix and process value chain analysis, transforming manual experience into a quantifiable feature matching model. Automatically replacing redundant nodes generates a lightweight process, overcoming the synergy gap between manual and automated processes. Finally, step S6 solidifies the entire lifecycle of multimodal data processing into a tamper-proof evidence chain through multi-level blockchain evidence storage and dual timestamp signatures, forming a closed-loop technology. This creates synergies across cross-modal analysis, dynamic rule adaptation, human-machine collaborative optimization, and trusted evidence storage. Through the deep coupling of data and control flows, it addresses the accuracy, real-time, and credibility issues in the open authentication of multimodal archives.

[0128] The calculations involved in the embodiments are all dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to actual conditions.

[0129] It should be noted that the present invention can be deployed on the device itself to realize embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0130] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0131] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0133] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0134] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0135] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0136] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0137] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for intelligent open identification of archives based on a large model, characterized in that: The steps include: S1. Perform cross-modal semantic analysis on the text data, image data and metadata of the archives, and generate multimodal feature vectors based on the hybrid expert model; S2. dynamically assigning the archival data to at least one processing path among a rule network for performing risk keyword matching, a semantic network for identifying cross-modal sensitive associations, and a domain network for matching regional features based on the attention weight distribution of each data modality in the multimodal feature vector; S3. Based on the regulatory update directive, the weight parameter optimization method based on gradient projection constraints is used to locally adjust the weight parameters related to risk keyword matching in the rule network; S4. Construct a multi-dimensional professional feature matrix for auditors, perform similarity matching between the cross-modal sensitive correlation data identified by the semantic network and the multi-dimensional professional feature matrix, and assign them to the audit nodes corresponding to the qualifications; S5. Perform value stream analysis on the audit process assigned to the audit node, remove manual confirmation steps and paper transmission nodes, and generate a lightweight audit process; S6. Write the multimodal feature vector, the dynamically allocated processing path, and the operation record of the lightweight audit process into the blockchain to generate a timestamped evidence chain.

2. The method for intelligent open identification of archives based on a large model according to claim 1, characterized in that: S1 includes: Perform cross-modal alignment on the text data, image data, and metadata of the archive, calculate the cross-modal semantic correlation between the text and image, and generate a cross-modal alignment feature vector; The cross-modal aligned feature vectors are input into the text expert network, image expert network, and metadata expert network in the hybrid expert model to extract text deep semantic features, image spatial features, and metadata spatiotemporal features, respectively. Perform multimodal feature fusion on text deep semantic features, image spatial features, and metadata spatiotemporal features, and calculate the weight coefficients of different modal features through a multi-head attention mechanism; The text deep semantic features, image spatial features and metadata spatiotemporal features are weighted and summed based on the weight coefficients to generate a multimodal feature vector.

3. The method for intelligent open identification of archives based on a large model according to claim 1, characterized in that S2 include: Normalize the attention weight distribution of text, image and metadata in the multimodal feature vector to generate text weight value, image weight value and metadata weight value; Generate at least one of a rule network trigger flag, a semantic network trigger flag, and a domain network trigger flag based on a comparison result of the text weight value and a preset text threshold, a comparison result of the image weight value and a preset image threshold, and a comparison result of the metadata weight value and a preset metadata threshold; Based on the state combination of the rule network trigger flag, the semantic network trigger flag and the domain network trigger flag, a dynamic processing path sequence is generated in the order that the rule network has a higher priority than the semantic network, and the semantic network has a higher priority than the domain network. The execution result of each processing path in the dynamic processing path sequence is used as the input condition for the next processing path.

4. The method for intelligent open identification of archives based on a large model according to claim 3 is characterized in that: When the rule network trigger flag is valid, the archive text data is filtered layer by layer based on the risk keyword matching rules in the rule network. The risk keyword matching rules include keyword whitelist, keyword blacklist and regular expression combination matching; When the semantic network trigger flag is valid, the implicit semantic association between text and image is extracted through the cross-modal association model, and the cross-modal sensitive information transmission path is identified based on the association strength threshold; When the domain network trigger flag is valid, a spatial overlap analysis is performed based on the geographic coordinate field in the archive metadata and the list of sensitive regions in the regional feature database to calculate the regional matching confidence.

5. The method for intelligent open identification of archives based on a large model according to claim 1, characterized in that: S3 includes: Analyze the new and abolished risk keywords in the regulatory update directives and generate a keyword update list; Weight parameters of the risk keyword matching rules associated with the keyword update list positioning rule network, the weight parameters include priority coefficient and matching sensitivity; Construct a weight parameter optimization model with gradient projection constraints. The constraints include the upper limit of the weight parameter variation and compliance verification rules. The positioning weight parameters are iteratively optimized through the gradient projection algorithm. The iterative termination condition is that the parameter change is less than the convergence threshold or the maximum number of iterations is reached; Update the optimized weight parameters to the rule network, perform validation set testing on the updated rule network, and activate the updated rule network after passing the test; When the validation set test fails, the system rolls back to the previous valid weight parameters and triggers the manual intervention process, which includes manual calibration of weight parameters and analysis of abnormal logs.

6. The method for intelligent open identification of archives based on a large model according to claim 1, characterized in that S4 include: Extract the auditor's professional field labels, historical audit records, and performance scores to generate the initial eigenvectors of the multi-dimensional professional feature matrix; Normalize the initial feature vectors to eliminate the dimensional differences of different feature dimensions and generate a normalized multi-dimensional professional feature matrix; Encode the cross-modal sensitive association data identified by the semantic network into sensitive feature vectors; Calculate the cosine similarity between the sensitive feature vector and each feature vector in the normalized multi-dimensional professional feature matrix to generate a similarity matching result; Based on the similarity matching results and the preset qualification threshold, the file data is dynamically allocated to the review node of the corresponding qualification; When the similarity matching result is lower than the qualification threshold, the cross-node collaborative review process is triggered. In the collaborative review process, the main review node and the auxiliary node jointly generate the review conclusion according to the weight ratio.

7. The method for intelligent open identification of archives based on a large model according to claim 6, characterized in that: The sensitive feature vector contains text sensitive entities, image sensitive areas and association strength values.

8. The method for intelligent open identification of archives based on a large model according to claim 1, characterized in that S5 include: Collect audit process data assigned to audit nodes, including task processing time, manual confirmation node identification, and paper transmission path; Identify manual confirmation links and paper transmission paths in the audit process as redundant nodes based on value stream analysis; Build a lightweight audit process optimization model, replace manual confirmation links with automated rule verification nodes, and replace paper transmission paths with blockchain digital signature transmission channels; Verify the integrity of the optimized process through process simulation. Verification conditions include meeting the task processing time compression ratio and passing the data hash consistency check. The lightweight audit process that has passed verification will be released in a phased manner, and the original process will be replaced in stages according to the load weight of the audit node; When a process exception occurs during a grayscale release, a rollback mechanism is triggered and an optimization defect report is generated.

9. The method for intelligent open identification of archives based on a large model according to claim 8, characterized in that: The defect report contains the exception type, triggering conditions, and repair suggestions.

10. The method for intelligent open identification of archives based on a large model according to claim 1, characterized in that S6 include: Perform data segmentation processing on multimodal feature vectors, dynamically allocated processing paths, and lightweight audit process operation records to generate a structured evidence data set; The structured evidence data set is split into multiple data blocks according to the preset block partitioning strategy, and each data block is attached with the hash value of the previous block to generate a chain association; Data blocks are written to the distributed ledger in parallel through blockchain nodes, and smart contracts are called during writing to verify data integrity and permission compliance; Generate a Merkle tree root hash value containing a write timestamp on the blockchain network and use the Merkle tree root hash value as the evidence chain identifier; A secondary timestamp signature is performed on the evidence chain identifier based on the timestamp service to generate an unalterable timestamp evidence chain. When the stored data needs to be verified, the data integrity and timestamp authenticity are verified by backtracing the Merkle tree path and matching the hash value.

Images