Bastion host sensitive data identification method based on large model
By dynamically adjusting the interception threshold and building a dynamic knowledge graph, the problem that traditional models cannot adapt to changes in complex scenarios is solved, and accurate identification and stable classification of bastion sensitive data is realized, which improves recognition efficiency and accuracy.
Patent Information
- Application Number
- CN202510563944.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-08-29
AI Technical Summary
In the prior art, traditional models cannot flexibly respond to changes in complex scenarios using fixed interception thresholds, lack real-time feedback and update mechanisms, resulting in insufficient classification accuracy, and the optimization of single attention indicators leads to unstable identification quality.
The bastion machine sensitive data recognition method based on large models is adopted, and the interception threshold is dynamically adjusted, and the dynamic knowledge graph is constructed for real-time updates. The model is optimized by combining reinforcement learning algorithms, pruning and lightweight processing is achieved to achieve accurate classification of sensitive data.
Accurate classification of sensitive data is realized, the adaptability and learning ability of the model are enhanced, the recognition efficiency and robustness are improved, and the continuous optimization and recognition quality of the system are ensured.
Smart Images

Figure CN120561723A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to information security technology, and in particular to a method for identifying sensitive data of a bastion host based on a large model. Background Art
[0002] Currently, traditional models typically use a fixed interception threshold. This approach often fails to adapt effectively to different data types, limiting the accuracy of classification results. Fixed thresholds lack the flexibility to adapt to complex scenarios, ultimately impacting overall system performance.
[0003] Secondly, many existing technologies lack real-time feedback and update mechanisms for models. Models struggle to absorb new recognition information during use, leading to long-term performance stagnation and an inability to adapt to changes in the data environment. As a result, models gradually become outdated, impacting their long-term effectiveness.
[0004] Furthermore, existing technologies often rely on manual intervention to update knowledge graphs. This is not only time-consuming and error-prone, but also inefficient in manually updating data and unable to meet rapidly changing needs. This lack of automated updating limits the real-time performance and accuracy of knowledge graphs.
[0005] Finally, many solutions tend to focus on a single metric, such as recall or precision, while ignoring the balance between the two. This one-sided optimization approach can lead to outstanding model performance in some scenarios, but significantly compromised performance in others, resulting in unstable recognition quality.
[0006] Therefore, in response to these deficiencies, the present invention arises as the times require, aiming to provide a more flexible and accurate sensitive data identification solution. Summary of the Invention
[0007] The purpose of this application is to provide a bastion host sensitive data identification method based on a large model, aiming to solve many deficiencies in the existing technology. The method achieves accurate classification of sensitive data by dynamically adjusting the interception threshold.
[0008] The present application provides a method for identifying sensitive data on a bastion host based on a large model, which adopts the following technical solution: S1: Preprocess the unstructured data collected by the bastion host, extract valid text information, generate corresponding word vector representations, and pre-train the large model; S2: After preprocessing is completed, a dynamic knowledge graph is constructed to store and manage sensitive entities and their relationships. The dynamic knowledge graph is updated based on new data monitored in real time. S3: Based on the sensitive data features in the dynamic knowledge graph, the pre-trained large model is pruned and lightweighted. S4: After the pruned model infers the input data, the entity features extracted from the dynamic knowledge graph are fused with the semantic features output by the model to identify and classify the sensitive data. S5: After the recognition and classification are completed, the interception threshold is dynamically adjusted according to the classification results, and the recognition result feedback is used to update the dynamic knowledge graph and large model parameters.
[0009] Preferably, the step of constructing a dynamic knowledge graph includes updating the weights of entities in the knowledge graph by the following formula: in: Represents entity e at time point t+1 i Update weight of Represents the weight at time point t; r t Represents the reward value obtained based on co-occurrence and misjudgment feedback; α represents the learning rate and controls the magnitude of the update.
[0010] Preferably, the dynamic knowledge graph update process adopts a reinforcement learning algorithm, and a reward function is designed to encourage the improvement of the knowledge graph in the identification and relevance of sensitive entities. The reward function includes: r t =α·f 共现 (e i ,e j )+β·f 误判 (e i ); in: f 共现 (e i ,e j ) represents entity e i With sensitive data j The co-occurrence frequency of f 误判 (e i ) represents the entity e i The misjudgment penalty is used to measure its negative impact on the overall recognition; α represents the weight coefficient of co-occurrence; β represents the penalty weight coefficient for misjudgment.
[0011] Preferably, the step of pruning the pre-trained large model implements the selection of attention heads by the following formula: in: H keep represents the set of retained attention heads; h i represents the i-th attention head; represents the attention head h i In the sensitive data corpus The score on τ represents the set threshold that determines which attention heads are important.
[0012] Preferably, the sensitive data is classified by a bilinear network, and the classification probability is output by calculating the weighted sum of the fusion features. The calculation formula is: in: p(y|h fused ) represents the classification probability of a given fusion feature; h fused Represents the fused feature vector; W represents the classification weight matrix, which is involved in calculating the importance of features; b represents the bias term, which adjusts the output; σ represents the Sigmoid function and outputs the probability value.
[0013] Preferably, the dynamic adjustment interception threshold is calculated based on the real-time load condition, and the dynamic threshold calculation formula is: η t =η base ·exp(-μQ t ); in: η t Indicates the current adjusted threshold; η base represents the baseline threshold; μ represents the attenuation coefficient, which controls the flexibility of the threshold; Q t Indicates the real-time load indicator of the system.
[0014] Preferably, the feedback mechanism includes feeding back misjudgment results obtained in manual review to the dynamic knowledge graph and the large model.
[0015] Preferably, the pruned model can be optimized by retraining, and the retraining process is as follows: new =TrainModel(X train ,Y train ); in: M new Represents the new model after retraining; TrainModel is a function or method used to train the model; X train Represents the training data set, the type is a multidimensional array; Y train Indicates the true label of the training data, the type is array.
[0016] Preferably, the implementation of the dynamic knowledge graph is based on graph neural network technology.
[0017] Preferably, the optimization of the dynamic knowledge graph continuously adjusts the knowledge graph structure through the user's evaluation data on the recognition results. The calculation formula of the optimization process is: G updated =G current +ψ·ΔG; in: G updated Represents the updated knowledge graph; G current Represents the current knowledge graph; ψ represents the adjustment coefficient, which indicates the degree of updating; ΔG represents the structural change of the knowledge graph based on user feedback.
[0018] By adopting the above technical solutions, many problems caused by fixed thresholds, such as insufficient classification accuracy, model failure, inefficient knowledge graph updates, and model instability, have been solved.
[0019] In summary, this application includes at least one of the following beneficial technical effects: 1. The present invention adopts a technical solution of dynamically adjusting the interception threshold to achieve accurate classification of sensitive data. Compared with the existing fixed threshold technology, it can optimize the model in real time according to the actual classification results, solving the problem that fixed thresholds are difficult to handle complex and changeable data features.
[0020] 2. This invention enhances the system's learning capabilities by timely updating the knowledge graph and model parameters. Compared to traditional solutions, this method can quickly respond to new recognition information, making the model more adaptable and overcoming the drawbacks of models that become ineffective over time and experience a decrease in recognition accuracy.
[0021] 3. This invention provides real-time feedback on interception results, ensuring continuous system optimization. Compared to previous update processes that required manual intervention, this automated feedback mechanism significantly improves efficiency and solves the problem of insufficient recognition accuracy caused by manual update delays.
[0022] 4. The present invention improves the comprehensiveness of sensitive data identification by comprehensively utilizing the dynamic balance between recall rate and false positive rate. Compared with the existing solutions that focus on a single indicator, the comprehensive evaluation method of the present invention ensures the robustness of the model in various scenarios and overcomes the disadvantage of unstable recognition effect in the past. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 It is a schematic diagram of the overall structure of Example X of the present application. DETAILED DESCRIPTION
[0024] The following is combined with Figure 1 , further details of this application are given.
[0025] Example: A method for identifying sensitive data of bastion host based on large model, referring to Figure 1 ,include: S1: Preprocess the unstructured data collected by the bastion host, extract valid text information, generate corresponding word vector representations, and pre-train the large model; Specifically, first, a data cleaning model is established. This model is mainly used to remove invalid information and noise data from the original input data. The cleaning content includes special characters, HTML tags, redundant spaces, etc. In this process, a regular expression (Regex) algorithm is used to filter strings, preferably using the following regular expression: CleanedText=RegexFilter(RawText); in: RawText represents the original input text, which is a string.
[0026] CleanedText represents the cleaned text generated after filtering, and its type is string.
[0027] Then, word segmentation is performed. A word segmentation algorithm based on deep learning is used to segment the cleaned text. The preferred embodiment uses LSTM (Long Short-Term Memory Network) or Word2Vec model to form a vocabulary sequence. The word segmentation algorithm can output the following word sequence: TokenizedWords=Tokenization(CleanedText); in: TokenizedWords is a word sequence generated after word segmentation, and its type is a word list.
[0028] Next, we use word embedding technology to generate corresponding word vectors and convert TokenizedWords into high-dimensional vector representations. In this regard, the preferred methods are GloVe (global vector representation) or BERT (bidirectional encoder representation). The word vector generation process is represented as follows: V i =WordEmbedding(TokenizedWords[i]); in: V i Represents the vector representation of the i-th word, the type is vector, and the dimension is d.
[0029] Furthermore, the generated word vectors are processed by aggregation technology. In this process, it is preferred to apply a pooling algorithm, such as maximum pooling or average pooling, to form a sentence-level representation. The aggregation process can be expressed as: V sentence =Pooling(V1,V2,…,V n ); in: V sentence Represents the generated sentence vector; V i The vector representation of the i-th word, the type is vector, usually with dimension d; n represents the number of words in the sentence, and its type is integer; Pooling represents a function that performs aggregation operations.
[0030] The main function of this operation is to integrate a set of word vectors into a sentence vector of fixed dimension, thereby providing a unified semantic representation for the entire sentence.
[0031] Finally, the output of step S1 is a word vector representation generated based on the input high-quality text data. This representation will serve as input to the subsequent sensitive data recognition model to achieve higher-precision recognition results. In summary, step S1 ensures that through a comprehensive data preprocessing process, highly usable and information-rich input data is provided to support the smooth progress of subsequent recognition tasks.
[0032] The technical process presented in this implementation effectively cleans, segment, and embed unstructured data, forming the foundation of the overall sensitive data identification process and providing stable technical support for subsequent model training and sensitive data identification. The specific implementation of these steps not only improves the model's ability to identify sensitive data but also facilitates further system optimization and intelligent development.
[0033] S2: After preprocessing is completed, a dynamic knowledge graph is constructed to store and manage sensitive entities and their relationships. The dynamic knowledge graph is updated based on new data monitored in real time. Specifically, first, an entity recognition model is established. This model is responsible for extracting sensitive data entities from text data. The entities include personal information (such as name, phone number, email address, etc.) and company information (such as company name, address, etc.). To achieve this function, a named entity recognition (NER) algorithm is preferably used, such as a conditional random field (CRF)-based or deep learning LSTM model. The formula is: Entities=NER(CleanedText); in: Entities represents the sensitive entities extracted from the cleaned text, and its type is entity list; CleanedText represents the cleaned text generated in the previous step, and its type is string.
[0034] Subsequently, a dynamic knowledge graph is constructed. The extracted entities and their relationships are organized into a graph structure, where nodes represent entities and edges represent relationships between entities. The initial construction of the knowledge graph can be expressed as: G = (E, R); in: G represents the knowledge graph, which is a graph structure; E represents the entity set, which contains all extracted sensitive entities and is of entity list type; R represents the set of relationships between entities, and its type is a relationship list.
[0035] Each entity e i ∈E can be expressed as: e i =(id i ,name i ,type i ,w i ); in: id i Represents the unique identifier of the entity, the type is string; name i Indicates the name of the entity, the type is string.
[0036] type i Indicates entity type, such as individual, company, etc., and the type is string.
[0037] w i Indicates the initial weight of the entity, the type is a floating point number.
[0038] Next, we derive and establish relationships between entities. By analyzing the context in the text or using semantic similarity calculation technology, we set edges to link different entities. The relationships can be expressed as: Rij =Relation(e i ,e j ); in: R ij Represents entity e i With e j The relationship between,type is relationship description; e i represents the i-th entity; e j represents the j-th entity.
[0039] Furthermore, the update of the dynamic knowledge graph is based on real-time monitoring and user feedback. After acquiring new data, the update of entity weights can be expressed as follows: in: Represents entity e at time point t+1 i Update weight of Represents the weight at time point t; r t Represents the reward value obtained based on co-occurrence and misjudgment feedback; α represents the learning rate and controls the magnitude of the update.
[0040] Furthermore, the dynamic knowledge graph update process adopts a reinforcement learning algorithm to encourage the improvement of the knowledge graph in the recognition and relevance of sensitive entities by designing a reward function. The reward function includes: t =α·f 共现 (e i ,e j )+β·f 误判 (e i ); in: f 共现 (e i ,e j ) represents entity e i With sensitive data j The co-occurrence frequency of f 误判 (e i ) represents the entity e i The misjudgment penalty is used to measure its negative impact on the overall recognition; α represents the weight coefficient of co-occurrence; β represents the penalty weight coefficient for misjudgment.
[0041] Finally, the updated entities and their relationships are stored in the knowledge graph to ensure that it reflects the latest information. By building and updating the dynamic knowledge graph, the system can effectively identify and manage sensitive data, improving its performance under real-time monitoring conditions.
[0042] S3: Prune and lightweight the pre-trained large model based on the sensitive data features in the dynamic knowledge graph. Specifically, first, a pruning model is established. This model analyzes the structure of the pre-trained large model and identifies unnecessary or redundant parameters to achieve lightweighting. The basic idea of pruning is to decide whether to remove each neuron or connection based on its contribution. Pruning can be achieved using the following formula: M pruned =M·Mask; in: M is the original model parameter; M pruned is the parameter after pruning; Mask is a binary mask vector, where a value of 1 indicates retention and a value of 0 indicates removal.
[0043] Secondly, important attention heads are selected for pruning. This process determines which attention heads to keep by calculating the contribution of each attention head to the final task. The formula can be expressed as: H keep ={hi|AttentionScore(hi)>τ}; in: H keep represents the set of retained attention heads; h i represents the i-th attention head; AttentionScore(h i ) represents the attention head h i The score of , which indicates its contribution to the task; τ represents the specified threshold, which is a floating point number and is set according to the importance of the task.
[0044] Next, the model is compressed. Quantization and sparsification techniques are preferably used to further reduce the storage and computational overhead of the model. Quantization compresses the storage space by converting floating-point parameters into low-precision integer representations. This process can be achieved using the following formula: w quantized =Quantize(w); in: w represents the floating-point model weight, the type is a floating-point array; w quantized Represents the quantized integer weight, the type is integer array.
[0045] Furthermore, the step of pruning the pre-trained large model implements the selection of attention heads through the following formula: keep ={h i |AttentionScore(h i )>τ}; in: H keep represents the set of retained attention heads; h i represents the i-th attention head; represents the attention head h i In the sensitive data corpus The score on τ represents the set threshold that determines which attention heads are important.
[0046] Through pruning and lightweighting, the model's reasoning efficiency can be significantly improved while maintaining recognition accuracy. This step ensures that unnecessary parameters of the model are effectively eliminated, thereby optimizing the overall system resource usage.
[0047] S4: After the pruned model infers the input data, the entity features extracted from the dynamic knowledge graph are fused with the semantic features output by the model to identify and classify sensitive data. Specifically, first, an evaluation model is established. This model is used to quantitatively evaluate the recognition results using standard evaluation metrics, including accuracy, recall, precision, and F1 score. Each metric can be calculated using the following formula: in: TP represents the true positives, the number of sensitive data correctly identified by the model; TN represents true negative examples, the number of data correctly identified as non-sensitive by the model; FP represents false positives, the number of data that the model incorrectly identifies as sensitive data; FN represents false negatives, the number of data that the model fails to identify as sensitive; Precision indicates the accuracy rate; Recall represents the recall rate.
[0048] Next, the model is evaluated using the test set. The test set includes labeled sensitive data samples and is used to compare the model's predictions with the true labels. The evaluation process is as follows: Predictions = ModelPredict(X test ,Mpruned ,w quantized ); in: Predictions: represents the output of the model for the test data. The type is an array, where each element represents the predicted label of the corresponding input data; X test Represents the test data set, which is a multidimensional array and contains the sample data to be evaluated; ModelPredict is a function or method responsible for performing the inference process of the model; M pruned Refers to the model parameters after pruning, including optimized weights and structures; w quantized Represents the quantized model weights, which are quantized to save storage space and speed up inference.
[0049] During this process, the output prediction results are compared with the true labels. Based on the true labels of each test sample, TP, TN, FP, and FN are calculated to obtain the above evaluation indicators.
[0050] Then, we optimize the results. Based on the evaluation results, we evaluate the model's performance in identifying sensitive data and develop an optimization strategy. The optimization steps include: Threshold adjustment: By analyzing the relationship between precision and recall, the ROC curve (receiver operating characteristic curve) is used to select the optimal threshold to balance the sensitivity and specificity of the model. The optimal threshold selected can be expressed as: τ=argmax τ (Recall-FPR); in: τ represents the decision threshold, which is used to determine the classification criteria of the prediction results; argmax is an operator that finds the parameter that makes the expression in the parentheses reach the maximum value; Recall stands for recall rate, which indicates the proportion of samples that the model correctly identifies among all samples that are actually sensitive data; FPR stands for false positive rate, which indicates the proportion of non-sensitive samples that the model mistakenly identifies as sensitive data.
[0051] Furthermore, the step of classifying sensitive data using a bilinear network outputs the classification probability by calculating the weighted sum of the fused features. The calculation formula is: in: p(y|h fused ) represents the classification probability of a given fusion feature; h fusedRepresents the fused feature vector; W represents the classification weight matrix, which is involved in calculating the importance of features; b represents the bias term, which adjusts the output; σ represents the Sigmoid function and outputs the probability value.
[0052] Finally, the optimized pruned model and quantized model are used to identify sensitive data. The reasoning steps in this process can be expressed as: Y=Inference(X,M pruned ,w quantized ); in: Y represents the output of the model, usually the identified sensitive data category or label. This variable is an array and contains the model's predictions for the input data. X represents the input data matrix, which is a multidimensional array. This data is a high-dimensional feature representation obtained through preprocessing and is used in the model's inference process. M pruned Represents the model parameters after pruning, as a multidimensional array. This parameter contains the optimized neural network weights, removing redundant and unnecessary connections to improve the model's inference efficiency; w quantized Represents the quantized model weights, as an integer array. These weights are quantized to reduce storage requirements while ensuring sufficient recognition accuracy during inference.
[0053] Model retraining: If the evaluation results show that the model is not performing well, the model is retrained using a new, more richly annotated training set. The retraining process is as follows: M new =TrainModel(X train ,Y train ); in: M new Represents the new model after retraining; TrainModel is a function or method used to train the model; X train Represents the training data set, the type is a multidimensional array; Y train Indicates the true label of the training data, the type is array.
[0054] Model integration: When necessary, the prediction results of multiple models can be combined to obtain the final recognition result through voting or weighted averaging to further improve accuracy.
[0055] After the optimization is completed, the model evaluation report is re-output to show the various indicators of the optimized model to ensure that it meets the actual application requirements. The report content should include: Final accuracy, recall, precision, and F1 score; Summary of key thresholds and optimization strategies used for optimization; Error analysis of test samples points out the types of samples in which the system is prone to errors, so as to facilitate subsequent improvements.
[0056] S5: After the recognition and classification is completed, the interception threshold is dynamically adjusted according to the classification results, and the recognition results are fed back to update the dynamic knowledge graph and large model parameters; Specifically, first, a dynamic interception threshold adjustment mechanism is established. This mechanism is used to adjust the interception threshold of the model in real time according to the classification results to achieve the optimal recognition effect. The adjustment strategy of the interception threshold τ is set as: τ new =τ old +α·(Recall-FPR); in: τ new represents the updated interception threshold; τ old Indicates the current interception threshold; α represents the adjustment coefficient, which is a positive value and is used to control the amplitude of the threshold adjustment; Recall represents the recall rate of the current model; FPR stands for false positive rate.
[0057] This adjustment mechanism ensures that the model can dynamically optimize its interval based on actual conditions when identifying sensitive data, thereby improving performance.
[0058] Furthermore, the step of dynamically adjusting the interception threshold value calculates a dynamic threshold value based on the real-time load condition. The dynamic threshold value calculation formula is: η t =η base ·exp(-μQ t ); in: η t Indicates the current adjusted threshold; η base represents the baseline threshold; μ represents the attenuation coefficient, which controls the flexibility of the threshold; Q t Indicates the real-time load indicator of the system.
[0059] Subsequently, through the annotated recognition results, the feedback information will be used to update the knowledge graph and optimize the model parameters. The specific process is as follows: First, samples corresponding to TP, FP, TN, and FN are extracted from the recognition results. These samples will serve as the basis for updating nodes and relationships in the knowledge graph. Secondly, adjust the knowledge graph based on the recognition results to ensure that the graph data is consistent with the newly recognized samples, and implement the following changes: UpdateGraph(TP,FP,TN,FN); in: TP, FP, TN, and FN represent the corresponding sample sets. The updating process enables the knowledge graph to reflect the latest learning results of the model.
[0060] Next, optimize the large model parameters. Based on the feedback results and the updated knowledge graph, use the following formula to retrain the model: M new =TrainModel(X train ,Y train ); in: M new Represents the updated model X train Indicates the updated sample feature dataset; Y train Represents the corresponding label dataset.
[0061] By retraining the model, new learning results will be fed back into the model, thereby improving its ability to identify sensitive data.
[0062] The design of the knowledge graph is optimized through the user feedback loop, and the user's evaluation data on the recognition results is used to continuously adjust the knowledge graph structure to make it more focused on high-frequency sensitive entities. The calculation formula of the optimization process is: G updated =G current +ψ·ΔG; in: G updated Represents the updated knowledge graph; G current Represents the current knowledge graph; ψ represents the adjustment coefficient, which indicates the degree of updating; ΔG represents the structural change of the knowledge graph based on user feedback.
[0063] Finally, through the process of dynamically adjusting the interception threshold, updating the knowledge graph, and retraining the model, the system can achieve continuous self-optimization and improvement.
[0064] The examples of this specific embodiment are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, any equivalent changes made based on the structure, shape, and principle of this application should be included in the scope of protection of this application.
Claims
1. A method for identifying sensitive data of a bastion host based on a large model, characterized in that: The following steps are involved: S1: Preprocess the unstructured data collected by the bastion host, extract valid text information, generate corresponding word vector representations, and pre-train the large model; S2: After preprocessing is completed, a dynamic knowledge graph is constructed to store and manage sensitive entities and their relationships. The dynamic knowledge graph is updated based on new data monitored in real time. S3: Prune and lightweight the pre-trained large model based on the sensitive data features in the dynamic knowledge graph. S4: After the pruned model infers the input data, the entity features extracted from the dynamic knowledge graph are fused with the semantic features output by the model to identify and classify sensitive data. S5: After the recognition and classification are completed, the interception threshold is dynamically adjusted according to the classification results, and the recognition result feedback is used to update the dynamic knowledge graph and large model parameters.
2. A method for identifying sensitive data of a bastion host based on a large model according to claim 1, characterized in that: The dynamic knowledge graph is constructed to update the weights of entities in the knowledge graph using the following formula: in: Represents entity e at time point t+1 i Update weight of Represents the weight at time point t; r t Represents the reward value obtained based on co-occurrence and misjudgment feedback; α represents the learning rate and controls the magnitude of the update.
3. A method for identifying sensitive data of a bastion host based on a large model according to claim 1, characterized in that: The dynamic knowledge graph update process uses a reinforcement learning algorithm to encourage the improvement of the knowledge graph in the identification and relevance of sensitive entities by designing a reward function. The reward function includes: r t =α·f 共现 (e i ,e j )+β·f 误判 (e i ); in: f 共现 (e i ,e j ) represents entity e i With sensitive data j The co-occurrence frequency of f 误判 (e i ) represents the entity e i The misjudgment penalty is used to measure its negative impact on the overall recognition; α represents the weight coefficient of co-occurrence; β represents the penalty weight coefficient for misjudgment.
4. A method for identifying sensitive data of a bastion host based on a large model according to claim 1, characterized in that: The pruning of the pre-trained large model is achieved by selecting the attention head through the following formula: in: H keep represents the set of retained attention heads; h i represents the i-th attention head; represents the attention head h i In the sensitive data corpus The score on τ represents the set threshold that determines which attention heads are important.
5. The method for identifying sensitive data of a bastion host based on a large model according to claim 1 is characterized in that: The sensitive data is classified through a bilinear network, and the classification probability is output by calculating the weighted sum of the fusion features. The calculation formula is: in: p(y|h fused ) represents the classification probability of a given fusion feature; h fused Represents the fused feature vector; W represents the classification weight matrix, which is involved in calculating the importance of features; b represents the bias term, which adjusts the output; σ represents the Sigmoid function and outputs the probability value.
6. A method for identifying sensitive data of a bastion host based on a large model according to claim 1, characterized in that: The dynamic adjustment interception threshold is calculated based on the real-time load condition. The dynamic threshold calculation formula is: or t =the base ·exp(-μQ t ); in: η t Indicates the current adjusted threshold; η base represents the baseline threshold; μ represents the attenuation coefficient, which controls the flexibility of the threshold; Q t Indicates the real-time load indicator of the system.
7. The method for identifying sensitive data of a bastion host based on a large model according to claim 1, characterized in that: The feedback mechanism includes feeding back the misjudgment results obtained in manual review to the dynamic knowledge graph and the large model.
8. The method for identifying sensitive data of a bastion host based on a large model according to claim 1, characterized in that: The pruned model can be optimized by retraining. The retraining process is as follows: M new =TrainModel(X train ,Y train ); in: M new Represents the new model after retraining; TrainModel is a function or method used to train the model; X train Represents the training data set, the type is a multidimensional array; Y train Indicates the true label of the training data, the type is array.
9. The method for identifying sensitive data of a bastion host based on a large model according to claim 1, characterized in that: The implementation of the dynamic knowledge graph is based on graph neural network technology.
10. The method for identifying sensitive data of a bastion host based on a large model according to claim 1, characterized in that: The optimization of the dynamic knowledge graph continuously adjusts the knowledge graph structure through the user's evaluation data on the recognition results. The calculation formula of the optimization process is: G updated =G current +ψ·ΔG; in: G updated Represents the updated knowledge graph; G current Represents the current knowledge graph; ψ represents the adjustment coefficient, which indicates the degree of updating; ΔG represents the structural change of the knowledge graph based on user feedback.
Citation Information
Cited By
Efficient sensitive data detection method
CN121561540A
An efficient sensitive data detection method
CN121561540B
