Large model data intelligent management method and system

Abstract

The invention discloses a large model data intelligent management method and system, and relates to the technical field of artificial intelligence data management, and the method comprises five steps of data collection, adversarial data detection, adversarial sample repair, data value evaluation and dynamic enhancement management. By introducing a bimodal detection mechanism which resists disturbance perception, combined recognition of structured numerical data and unstructured text data is realized; the repair security and semantic consistency of the adversarial sample are improved through a repair strategy combining disturbance backstepping and semantic reconstruction; dynamic quantification and screening of data values are realized in combination with a quantitative evaluation method based on feature contribution degree and information entropy; according to the method, the quality control capability and the resource utilization efficiency of large model training data can be effectively improved, and the method has relatively high practicability and innovativeness.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence data management technology, and specifically to a large-scale model data intelligent management method and system. Background Art

[0002] The large-scale model data intelligent management method and system is a technical solution for efficiently and securely managing and optimizing large-scale training data. It aims to improve data quality and utilization efficiency through automated data collection, anomaly detection, sample repair and value assessment, and ensure the stability, robustness and performance of large artificial intelligence models during training and application.

[0003] However, in the existing large-model data intelligent management methods, there are three technical problems that are difficult to solve collaboratively: high data heterogeneity, strong concealment of adversarial pollution, and low data utilization efficiency; in the existing data adversarial detection process, there are technical problems such as the separation of numerical and textual modes, the reliance of detection methods on static rules, and the lack of model-level response measurements to perturbation behaviors; in the existing adversarial data sample repair process, there are technical problems such as the reliance of numerical sample repair on static distribution, the lack of semantic fidelity control in text repair, and the poor interpretability of the overall repair, making it difficult to trace the source. Summary of the Invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a large-model data intelligent management method and system. In view of the three technical problems that are difficult to solve collaboratively in the existing large-model data intelligent management methods, namely, high data heterogeneity, strong anti-pollution concealment and low data utilization efficiency, this solution creatively adopts an integrated data management method that combines adversarial detection, adversarial repair and comprehensive evaluation of data value, so as to achieve dynamic improvement of the quality controllability and utilization priority of large-model training data while ensuring data security, and significantly enhance the support capability of basic data for model performance; in view of the technical problems that exist in the existing data adversarial detection process, such as the separation of numerical and text modes, the reliance of detection methods on static rules, and the lack of model-level response measurement for disturbance behavior, this solution creatively adopts dual-modal data combined with adversarial disturbance perception. The detection method conducts adversarial data detection, realizes the joint detection and responsive recognition of structured numerical data and unstructured text data, and uniformly outputs the detection results through the confidence fusion mechanism, thereby improving the consistency and controllability of adversarial sample identification; in view of the technical problems in the repair process of existing adversarial data samples, there are numerical sample repairs that rely on static distribution, text repairs lack semantic fidelity control, and the overall repair has poor interpretability and is difficult to trace back and check, this scheme creatively adopts an adversarial sample repair method that combines perturbation backpropagation and semantic reconstruction to repair adversarial samples, achieving security enhancement and semantic consistency restoration of samples in sensitive areas of model output. The numerical part realizes accurate callback of the input value through perturbation direction backpropagation, and the text part combines keyword normalization replacement with fixed sentence rewriting to ensure the support of the repaired samples for the downstream training process.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides a large model data intelligent management method, the method comprising the following steps:

[0006] Step S1: data collection;

[0007] Step S2: adversarial data detection;

[0008] Step S3: adversarial sample repair;

[0009] Step S4: Data value assessment;

[0010] Step S5: Dynamic enhancement of large model data management.

[0011] Furthermore, in step S1, the data collection is used to collect the original data set of the large model, specifically from the behavior log data, the customer rating system and the interface call monitoring system log summary, through data collection, to obtain the original data of the large model, and through data preprocessing operations, to obtain the structured original data of the large model;

[0012] The large model original data includes platform logs, interaction records and feedback corpus data.

[0013] Furthermore, in step S2, the adversarial data detection is used to identify adversarial numerical anomalies on the original data. Specifically, based on the original data of the structured large model, a bimodal data detection method combined with adversarial disturbance perception is used to perform adversarial data detection to obtain adversarial detection enhanced data, including the following steps:

[0014] Step S21: numerical abnormal disturbance detection, specifically applying data disturbance to the numerical data sample, and constructing a standard isolation forest model combined with dynamic threshold adjustment to perform numerical abnormal disturbance detection, and obtain disturbance detection numerical data;

[0015] The calculation formula of the standard isolation forest model is:

[0016] ;

[0017] Where s i is the anomaly score of the numerical sample, when the anomaly score s i When it is greater than the default threshold, it is marked as an adversarial sample, i is the numerical sample index, E(·) is the standard isolation tree average path length calculation function, h i is the identifier of the i-th numerical sample, c(k) is the normalization factor when the total number of samples is k, and H(k) is the harmonic number parameter. The specific calculation formula is , k is the total number of numerical samples;

[0018] The calculation formula for the dynamic threshold adjustment is:

[0019] ;

[0020] Where, is the numerical abnormality judgment threshold at the t+1th iteration, is the numerical abnormality judgment threshold at the tth iteration, is the learning rate parameter, FPR target is the expected probability target value of judging that the adversarial sample is true, FPR actual It is the true value of the probability of judging that the adversarial sample is true;

[0021] Step S22: Text adversarial sample detection, specifically constructing a dual-branch text detection mechanism, including gradient sensitivity detection and sentiment word matching detection, performing text adversarial sample detection, and obtaining text adversarial sample detection data;

[0022] The gradient sensitivity detection is specifically performed by calculating the gradient sensitivity score;

[0023] The sentiment word matching detection specifically performs semantic similarity detection through a predefined sentiment word template library;

[0024] The predefined sentiment word template library includes a positive-negative polarity replacement word library, a neutral-negative polarity replacement word library, and a negative-positive polarity replacement word library;

[0025] Step S23: Confidence fusion, specifically, performing joint probability calculation based on the disturbance detection numerical data and the text adversarial sample detection data using a joint probability synthesis method to obtain fused adversarial probability reference data, and performing adversarial data result detection based on the fused adversarial probability reference data by setting a detection threshold;

[0026] The calculation formula for the joint probability calculation is:

[0027] ;

[0028] Where, P adv is the comprehensive confrontation probability value, is the comprehensive weight, is the standard indicator function identifier, s i is the anomaly score of the numerical sample, is the threshold value for judging numerical anomaly, C(x text ) is a text adversarial judgment indicator, which specifically adopts the weighted combination of gradient sensitivity detection and sentiment word matching detection results;

[0029] Step S24: adversarial data detection, specifically, performing adversarial data detection through the numerical anomaly combination detection, the text adversarial sample detection and the confidence fusion to obtain adversarial detection enhanced data.

[0030] Furthermore, in step S3, the adversarial sample repair is used to convert the adversarial sample into safe and usable data by combining rules with the generative model. Specifically, based on the adversarial detection enhanced data, an adversarial sample repair method combining perturbation back-inference and semantic reconstruction is used to perform adversarial sample repair to obtain adversarial sample repair enhanced data, including the following steps:

[0031] Step S31: numerical disturbance back-calculation and repair, specifically, performing numerical disturbance back-calculation and repair according to the disturbance detection method of step S21, constructing disturbance repair data, and obtaining disturbance repair numerical data;

[0032] The calculation formula for the numerical perturbation back-stepping repair is:

[0033] ;

[0034] Where x repaired is the perturbation repair numerical data, x advis the numerical input identified as an adversarial example, is the perturbation amplitude coefficient, sign(·) is the standard element-wise sign function identifier, is the gradient of the loss function of the standard isolation forest model,

[0035] is the output of the standard isolation forest model, is the input data sample of the standard isolation forest model, is the output data sample of the standard isolation forest model;

[0036] Step S32: normalizing and replacing text keywords, specifically performing normalization and replacement of text keywords based on the predefined sentiment word template library in the text adversarial sample detection in step S22 to obtain text word replacement repair data;

[0037] Step S33: text semantic repair, specifically, performing sentence template semantic repair detection based on the text repair data, performing text semantic repair, and obtaining text sentence semantic repair data;

[0038] Step S34: repair and improve the adversarial sample, specifically repairing and improving the adversarial sample through the numerical perturbation inverse repair, the text keyword normalization replacement and the text semantic repair, to obtain adversarial sample repair enhancement data.

[0039] Furthermore, in step S4, the data value assessment is used to dynamically quantify the contribution of data to model training. Specifically, based on the adversarial detection enhancement data and the adversarial sample repair enhancement data, a comprehensive normalization improved numerical value analysis method is used to perform data value assessment to obtain data value contribution assessment reference data, including the following steps:

[0040] Step S41: numerical data value assessment, specifically, using a standard fitted linear regression model based on feature contribution inversion, and using the absolute value of the regression coefficient of each numerical data to the target variable as the prediction contribution score to perform numerical data value assessment and obtain numerical value assessment reference data;

[0041] Step S42: evaluating the value of text data, specifically using a standard text information entropy evaluation method based on the TF-IDF matrix to evaluate the value of text data and obtain reference data for text value evaluation;

[0042] Step S43: score normalization, specifically, by using a comprehensive normalization method to quantify and label the data value based on the numerical value evaluation reference data and the text value evaluation reference data, to obtain data value normalized reference data;

[0043] The comprehensive normalization method includes quantitative normalization and label normalization, and the calculation formula is:

[0044] ;

[0045] Where S n is the quantitative normalized value assessment data, V n is the original value evaluation data extracted from the numerical value evaluation reference data and the text value evaluation reference data, min(·) is the minimum value obtaining function, V is the complete set of evaluation value data composed of the numerical value evaluation reference data and the text value evaluation reference data, max(·) is the maximum value obtaining function, and L n It is the label normalized evaluation data calculated based on the quantitative normalized value evaluation data;

[0046] Step S44: Comprehensive evaluation of data value, specifically, performing comprehensive evaluation and normalization of data value through the numerical data value evaluation, the text data value evaluation and the score normalization to obtain reference data for data value contribution evaluation.

[0047] Furthermore, in step S5, the large model data management is dynamically enhanced to combine adversarial detection, repair and data value assessment to optimize the overall utilization of large model data resources. Specifically, the adversarial detection enhanced data and the adversarial sample repair enhanced data are combined to integrate the large model data source data to obtain fused enhanced structured data, and the data value is arranged according to the data value contribution assessment reference data. By setting a dynamic threshold for value contribution, data is screened to obtain dynamically enhanced large model basic data.

[0048] The present invention provides a large-model data intelligent management system, which includes a data collection module, an adversarial data detection module, an adversarial sample repair module, a data value assessment module, and a large-model data management dynamic enhancement module;

[0049] The data collection module is used for data collection, obtains the original data of the structured large model through data collection, and sends the original data of the structured large model to the adversarial data detection module;

[0050] The adversarial data detection module is used for adversarial data detection, obtains adversarial detection enhancement data through adversarial data detection, and sends the adversarial detection enhancement data to the adversarial sample repair module and the data value assessment module;

[0051] The adversarial sample repair module is used to repair the adversarial sample, obtain adversarial sample repair enhancement data through adversarial sample repair, and send the adversarial sample repair enhancement data to the data value assessment module;

[0052] The data value assessment module is used for data value assessment, obtains data value contribution assessment reference data through data value assessment, and sends the data value contribution assessment reference data to the large model data management dynamic enhancement module;

[0053] The large model data management dynamic enhancement module is used for dynamic enhancement of large model data management, and obtains dynamically enhanced large model basic data through dynamic enhancement of large model data management.

[0054] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0055] (1) In view of the three technical problems that are difficult to solve collaboratively in existing large-scale model data intelligent management methods, namely high data heterogeneity, strong adversarial contamination concealment, and low data utilization efficiency, this solution creatively adopts an integrated data management method that combines adversarial detection, adversarial repair, and comprehensive data value evaluation. While ensuring data security, it dynamically improves the quality controllability and utilization priority of large-scale model training data, significantly enhancing the support capability of basic data for model performance.

[0056] (2) In response to the technical problems in the existing data adversarial detection process, such as the separation of numerical and textual modalities, the reliance on static rules for detection methods, and the lack of model-level response measurement for perturbation behavior, this solution creatively adopts a dual-modal data detection method combined with adversarial perturbation perception to perform adversarial data detection, achieving joint detection and responsive recognition of structured numerical data and unstructured textual data, and uniformly outputting detection results through a confidence fusion mechanism, thereby improving the consistency and controllability of adversarial sample recognition;

[0057] (3) In view of the technical problems in the repair process of existing adversarial data samples, such as the dependence of numerical sample repair on static distribution, the lack of semantic fidelity control in text repair, the poor interpretability of the overall repair, and the difficulty in tracing the source, this solution creatively adopts an adversarial sample repair method that combines perturbation back-inference and semantic reconstruction to repair adversarial samples, achieving security enhancement and semantic consistency restoration of samples in the sensitive area of model output. The numerical part realizes accurate callback of the input value by back-inference of the perturbation direction, and the text part combines keyword normalization replacement with fixed sentence rewriting to ensure the support of the repaired samples for the downstream training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A flow chart of a large model data intelligent management method provided by the present invention;

[0059] Figure 2 A schematic diagram of a large model data intelligent management system provided by the present invention;

[0060] Figure 3 This is a flow chart of the adversarial data detection process in step S2;

[0061] Figure 4 Schematic diagram of the process of adversarial sample repair in step S3;

[0062] Figure 5 Schematic diagram of the process of data value assessment in step S4.

[0063] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0065] In the description of the present invention, it should be understood that terms such as "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they should not be understood as limiting the present invention.

[0066] Example 1, see Figure 1 The present invention provides a large model data intelligent management method, which includes the following steps:

[0067] Step S1: data collection;

[0068] Step S2: adversarial data detection;

[0069] Step S3: adversarial sample repair;

[0070] Step S4: Data value assessment;

[0071] Step S5: Dynamic enhancement of large model data management.

[0072] By performing the above operations, this solution creatively adopts an integrated data management method that combines adversarial detection, adversarial repair and comprehensive data value evaluation to address the three technical problems that are difficult to solve collaboratively in existing large-model data intelligent management methods: high data heterogeneity, strong adversarial pollution concealment and low data utilization efficiency. It achieves the dynamic improvement of the quality controllability and utilization priority of large-model training data while ensuring data security, and significantly enhances the basic data's support for model performance.

[0073] Example 2, see Figure 1 and Figure 2 ,In step S1, the data collection is used to collect the original data set of the large model, specifically from the behavior log data, the customer rating system and the interface call monitoring system log summary, through data collection, the original data of the large model is obtained, and through data preprocessing operations, the structured original data of the large model is obtained;

[0074] The large model raw data includes platform logs, interaction records and feedback corpus data;

[0075] The data type of the large model original data includes numerical data and text data;

[0076] The data preprocessing operation includes numerical data preprocessing and text data preprocessing;

[0077] The numerical data preprocessing includes numerical type identification, anomaly filtering and data standardization;

[0078] The text data preprocessing includes text segmentation, sensitive word filtering, adversarial sample generation, encoding conversion, field labeling and index generation;

[0079] The structured large model original data specifically includes numerical feature data, text content data, type tag data, sample number data and time feature data.

[0080] Example 3, see Figure 1 、 Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the adversarial data detection is used to identify adversarial numerical anomalies on the original data. Specifically, based on the original data of the structured large model, a dual-modal data detection method combined with adversarial disturbance perception is used to perform adversarial data detection to obtain adversarial detection enhanced data, including the following steps:

[0081] Step S21: numerical abnormal disturbance detection, specifically applying data disturbance to the numerical data sample, and constructing a standard isolation forest model combined with dynamic threshold adjustment to perform numerical abnormal disturbance detection, and obtain disturbance detection numerical data;

[0082] The calculation formula of the standard isolation forest model is:

[0083] ;

[0084] Where s i is the anomaly score of the numerical sample, when the anomaly score s i When it is greater than the default threshold, it is marked as an adversarial sample, i is the numerical sample index, E(·) is the standard isolation tree average path length calculation function, h i is the identifier of the i-th numerical sample, c(k) is the normalization factor when the total number of samples is k, and H(k) is the harmonic number parameter. The specific calculation formula is , k is the total number of numerical samples;

[0085] Preferably, the default threshold value is 0.65 and is updated through dynamic threshold adjustment;

[0086] The calculation formula for the dynamic threshold adjustment is:

[0087] ;

[0088] Where, is the numerical abnormality judgment threshold at the t+1th iteration, is the numerical abnormality judgment threshold at the tth iteration, is the learning rate parameter, FPR target is the expected probability target value of judging that the adversarial sample is true, FPR actual It is the true value of the probability of judging that the adversarial sample is true;

[0089] Step S22: Text adversarial sample detection, specifically constructing a dual-branch text detection mechanism, including gradient sensitivity detection and sentiment word matching detection, performing text adversarial sample detection, and obtaining text adversarial sample detection data;

[0090] The gradient sensitivity test is specifically performed by calculating the gradient sensitivity score, and the calculation formula is:

[0091] ;

[0092] Where, is the gradient sensitivity score. When the gradient sensitivity score is greater than the default sensitivity threshold of the text, it is marked as an adversarial sample. is the gradient of the loss function, is the output of the text classification model, specifically the standard BERT model, y is the true label of the text sample, and x is the original text sample;

[0093] Preferably, the default sensitivity threshold of the text is 1.2;

[0094] The sentiment word matching detection specifically performs semantic similarity detection through a predefined sentiment word template library, and the calculation formula is:

[0095] ;

[0096] Where m(x) is the semantic similarity score of sentiment word matching, T is the predefined sentiment word template library, Is a predefined sentiment word index, Sim BERT (·) is the original text x and predefined sentiment words calculated using the standard BERT model cosine similarity between ;

[0097] The predefined sentiment word template library includes a positive-negative polarity replacement word library, a neutral-negative polarity replacement word library, and a negative-positive polarity replacement word library;

[0098] Step S23: Confidence fusion, specifically, performing joint probability calculation based on the disturbance detection numerical data and the text adversarial sample detection data using a joint probability synthesis method to obtain fused adversarial probability reference data, and performing adversarial data result detection based on the fused adversarial probability reference data by setting a detection threshold;

[0099] The calculation formula for the joint probability calculation is:

[0100] ;

[0101] Where, P adv is the comprehensive confrontation probability value, is the comprehensive weight, is the standard indicator function identifier, s i is the anomaly score of the numerical sample, is the threshold value for judging numerical anomaly, C(x text ) is a text adversarial judgment indicator, which specifically adopts the weighted combination of gradient sensitivity detection and sentiment word matching detection results;

[0102] Preferably, the comprehensive weight The specific value of is 0.45;

[0103] Step S24: adversarial data detection, specifically, performing adversarial data detection through the numerical anomaly combination detection, the text adversarial sample detection and the confidence fusion to obtain adversarial detection enhanced data.

[0104] By performing the above operations, this solution creatively adopts a dual-modal data detection method combined with adversarial perturbation perception to detect adversarial data, addressing the technical problems in the existing data adversarial detection process, such as the separation of numerical and textual modalities, the reliance of detection methods on static rules, and the lack of model-level response measurement to perturbation behaviors. It realizes the joint detection and responsive recognition of structured numerical data and unstructured textual data, and uniformly outputs the detection results through a confidence fusion mechanism, thereby improving the consistency and controllability of adversarial sample recognition.

[0105] Example 4, see Figure 1 、 Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S3, the adversarial sample repair is used to convert the adversarial sample into safe and usable data by combining rules with the generative model. Specifically, based on the adversarial detection enhanced data, an adversarial sample repair method combining perturbation back-inference and semantic reconstruction is used to repair the adversarial sample to obtain the adversarial sample repair enhanced data, including the following steps:

[0106] Step S31: numerical disturbance back-calculation and repair, specifically, performing numerical disturbance back-calculation and repair according to the disturbance detection method of step S21, constructing disturbance repair data, and obtaining disturbance repair numerical data;

[0107] The calculation formula for the numerical perturbation back-stepping repair is:

[0108] ;

[0109] Where x repaired is the perturbation repair numerical data, x adv is the numerical input identified as an adversarial example, is the perturbation amplitude coefficient, sign(·) is the standard element-wise sign function identifier, is the gradient of the loss function of the standard isolation forest model,

[0110] is the output of the standard isolation forest model, is the input data sample of the standard isolation forest model, is the output data sample of the standard isolation forest model;

[0111] Step S32: normalizing and replacing text keywords, specifically performing normalization and replacement of text keywords based on the predefined sentiment word template library in the text adversarial sample detection in step S22 to obtain text word replacement repair data;

[0112] The calculation formula for normalized replacement of text keywords is:

[0113] ;

[0114] Where, It is the text after replacement normalization, used to represent the text word replacement repair data, is the replacement execution function, used to represent the text word replacement operation, T adv is the original adversarial text, M is the total number of keywords to be matched based on the predefined sentiment word template library, m is the index of the keyword to be matched, and w m is the replacement indicator variable, t m is the identified word to be replaced, It is a replacement word;

[0115] Step S33: text semantic repair, specifically, performing sentence template semantic repair detection based on the text repair data, performing text semantic repair, and obtaining text sentence semantic repair data;

[0116] The calculation formula for the sentence template semantic repair detection is:

[0117] ;

[0118] Where, is the text sentence semantic repair data, is the semantic repair data rewritten by the sentence template, T adv is the original adversarial text, Sim BERT (·) is the semantic restoration data calculated using the standard BERT model and the original adversarial text T adv Semantic similarity between

[0119] Step S34: repair and improve the adversarial sample, specifically repairing and improving the adversarial sample through the numerical perturbation inverse repair, the text keyword normalization replacement and the text semantic repair, to obtain adversarial sample repair enhancement data.

[0120] By performing the above operations, in order to address the technical problems in the repair process of existing adversarial data samples, such as the reliance of numerical sample repair on static distribution, the lack of semantic fidelity control in text repair, the poor interpretability of the overall repair, and the difficulty in tracing the source, this solution creatively adopts an adversarial sample repair method that combines perturbation backpropagation and semantic reconstruction to repair adversarial samples, thereby achieving security enhancement and semantic consistency restoration for samples in sensitive areas of model output. The numerical part achieves precise callback of the input value through perturbation direction backpropagation, and the text part combines keyword normalization replacement with fixed sentence rewriting to ensure that the repaired samples support the downstream training process.

[0121] Example 5, see Figure 1 、 Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the data value assessment is used to dynamically quantify the contribution of data to model training. Specifically, based on the adversarial detection enhancement data and the adversarial sample repair enhancement data, a comprehensive normalization improved numerical value analysis method is used to perform data value assessment to obtain data value contribution assessment reference data, including the following steps:

[0122] Step S41: numerical data value assessment, specifically, using a standard fitted linear regression model based on feature contribution inversion, and using the absolute value of the regression coefficient of each numerical data to the target variable as the prediction contribution score to perform numerical data value assessment and obtain numerical value assessment reference data;

[0123] Step S42: evaluating the value of text data, specifically using a standard text information entropy evaluation method based on the TF-IDF matrix to evaluate the value of text data and obtain reference data for text value evaluation;

[0124] Step S43: score normalization, specifically, by using a comprehensive normalization method to quantify and label the data value based on the numerical value evaluation reference data and the text value evaluation reference data, to obtain data value normalized reference data;

[0125] The comprehensive normalization method includes quantitative normalization and label normalization, and the calculation formula is:

[0126] ;

[0127] Where S n is the quantitative normalized value assessment data, V n is the original value evaluation data extracted from the numerical value evaluation reference data and the text value evaluation reference data, min(·) is the minimum value obtaining function, V is the complete set of evaluation value data composed of the numerical value evaluation reference data and the text value evaluation reference data, max(·) is the maximum value obtaining function, and L n It is the label normalized evaluation data calculated based on the quantitative normalized value evaluation data;

[0128] Step S44: Comprehensive evaluation of data value, specifically, performing comprehensive evaluation and normalization of data value through the numerical data value evaluation, the text data value evaluation and the score normalization to obtain reference data for data value contribution evaluation.

[0129] Example 6, see Figure 1 and Figure 2This embodiment is based on the above embodiment. In step S5, the large model data management is dynamically enhanced to combine adversarial detection, repair and data value evaluation to optimize the overall utilization of large model data resources. Specifically, the adversarial detection enhanced data and the adversarial sample repair enhanced data are combined to integrate the large model data source data to obtain fused enhanced structured data, and the data value is arranged according to the data value contribution evaluation reference data. By setting the dynamic threshold of value contribution, data is screened to obtain dynamically enhanced large model basic data.

[0130] Example 7, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. The present invention provides a large model data intelligent management system, including a data collection module, an adversarial data detection module, an adversarial sample repair module, a data value assessment module and a large model data management dynamic enhancement module;

[0131] The data collection module is used for data collection, obtains the original data of the structured large model through data collection, and sends the original data of the structured large model to the adversarial data detection module;

[0132] The adversarial data detection module is used for adversarial data detection, obtains adversarial detection enhancement data through adversarial data detection, and sends the adversarial detection enhancement data to the adversarial sample repair module and the data value assessment module;

[0133] The adversarial sample repair module is used to repair the adversarial sample, obtain adversarial sample repair enhancement data through adversarial sample repair, and send the adversarial sample repair enhancement data to the data value assessment module;

[0134] The data value assessment module is used for data value assessment, obtains data value contribution assessment reference data through data value assessment, and sends the data value contribution assessment reference data to the large model data management dynamic enhancement module;

[0135] The large model data management dynamic enhancement module is used for dynamic enhancement of large model data management, and obtains dynamically enhanced large model basic data through dynamic enhancement of large model data management.

[0136] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0137] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that various changes, modifications, substitutions, and alterations can be made to the embodiments without departing from the principles and spirit of the invention.

[0138] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A large-scale model data intelligent management method, characterized by: The method comprises the following steps: Step S1: Data collection, obtaining the original data of the structured large model; Step S2: adversarial data detection, using a dual-modal data detection method combined with adversarial perturbation perception to perform adversarial data detection and obtain adversarial detection enhanced data, including the following steps: Step S21: numerical abnormal perturbation detection, constructing a standard isolation forest model combined with dynamic threshold adjustment to perform numerical abnormal perturbation detection; Step S22: text adversarial sample detection; Step S23: confidence fusion, using a joint probability synthesis method to perform joint probability calculation; Step S24: adversarial data detection; Step S3: adversarial sample repair, using an adversarial sample repair method that combines perturbation back-inference and semantic reconstruction to perform adversarial sample repair and obtain adversarial sample repair enhancement data, including the following steps: step S31: numerical perturbation back-inference repair, performing numerical perturbation back-inference repair according to the perturbation detection method in step S21, and constructing perturbation repair data; step S32: text keyword normalization replacement; step S33: text semantic repair; step S34: adversarial sample repair improvement; Step S4: Data value assessment uses a comprehensive normalization improved numerical value analysis method to conduct data value assessment and obtain reference data for data value contribution assessment; Step S5: Dynamically enhance the large model data management to obtain dynamically enhanced large model basic data.

2. A large-scale model data intelligent management method according to claim 1, characterized in that: In step S1, the data collection is used to collect the original data set of the large model, specifically from the behavior log data, the customer rating system and the interface call monitoring system log summary, through data collection, to obtain the original data of the large model, and through data preprocessing operations, to obtain the structured original data of the large model; The large model raw data includes platform logs, interaction records and feedback corpus data; The data types of the large model original data include numerical data and text data.

3. The method for intelligent management of large model data according to claim 2, characterized in that: In step S2, the adversarial data detection is used to identify adversarial numerical anomalies on the original data. Specifically, based on the original data of the structured large model, a bimodal data detection method combined with adversarial disturbance perception is used to perform adversarial data detection to obtain adversarial detection enhanced data, including the following steps: Step S21: numerical abnormal disturbance detection, specifically applying data disturbance to the numerical data sample, and constructing a standard isolation forest model combined with dynamic threshold adjustment to perform numerical abnormal disturbance detection, and obtain disturbance detection numerical data; The calculation formula of the standard isolation forest model is: ； Where s i is the anomaly score of the numerical sample, when the anomaly score s i When it is greater than the default threshold, it is marked as an adversarial sample, i is the numerical sample index, E(·) is the standard isolation tree average path length calculation function, h i is the identifier of the i-th numerical sample, c(k) is the normalization factor when the total number of samples is k, and H(k) is the harmonic number parameter. The specific calculation formula is , k is the total number of numerical samples; The calculation formula for the dynamic threshold adjustment is: ； Where, is the numerical abnormality judgment threshold at the t+1th iteration, is the numerical abnormality judgment threshold at the tth iteration, is the learning rate parameter, FPR target is the expected probability target value of judging that the adversarial sample is true, FPR actual It is the true value of the probability of judging that the adversarial sample is true; Step S22: Text adversarial sample detection, specifically constructing a dual-branch text detection mechanism, including gradient sensitivity detection and sentiment word matching detection, performing text adversarial sample detection, and obtaining text adversarial sample detection data; The gradient sensitivity detection is specifically performed by calculating the gradient sensitivity score; The sentiment word matching detection specifically performs semantic similarity detection through a predefined sentiment word template library; The predefined sentiment word template library includes a positive-negative polarity replacement word library, a neutral-negative polarity replacement word library, and a negative-positive polarity replacement word library; Step S23: Confidence fusion, specifically, performing joint probability calculation based on the disturbance detection numerical data and the text adversarial sample detection data using a joint probability synthesis method to obtain fused adversarial probability reference data, and performing adversarial data result detection based on the fused adversarial probability reference data by setting a detection threshold; Step S24: adversarial data detection, specifically, performing adversarial data detection through the numerical anomaly combination detection, the text adversarial sample detection and the confidence fusion to obtain adversarial detection enhanced data.

4. A large-scale model data intelligent management method according to claim 3, characterized in that: In step S23, the calculation formula for the joint probability calculation is: ； Where, P adv is the comprehensive confrontation probability value, is the comprehensive weight, is the standard indicator function identifier, s i is the anomaly score of the numerical sample, is the threshold value for judging numerical anomaly, C(x text ) is a text adversarial judgment indicator, which specifically adopts the weighted combination of gradient sensitivity detection and sentiment word matching detection results.

5. The method for intelligent management of large model data according to claim 4, characterized in that: In step S3, the adversarial sample repair is used to convert the adversarial sample into safe and usable data by combining rules with the generative model. Specifically, based on the adversarial detection enhanced data, an adversarial sample repair method combining perturbation back-inference and semantic reconstruction is used to repair the adversarial sample to obtain the adversarial sample repair enhanced data, including the following steps: Step S31: numerical disturbance back-calculation and repair, specifically, performing numerical disturbance back-calculation and repair according to the disturbance detection method of step S21, constructing disturbance repair data, and obtaining disturbance repair numerical data; The calculation formula for the numerical perturbation back-stepping repair is: ； Where x repaired is the perturbation repair numerical data, x adv is the numerical input identified as an adversarial example, is the perturbation amplitude coefficient, sign(·) is the standard element-wise sign function identifier, is the gradient of the loss function of the standard isolation forest model, is the output of the standard isolation forest model, is the input data sample of the standard isolation forest model, is the output data sample of the standard isolation forest model; Step S32: normalizing and replacing text keywords, specifically performing normalization and replacement of text keywords based on the predefined sentiment word template library in the text adversarial sample detection in step S22 to obtain text word replacement repair data; Step S33: text semantic repair, specifically, performing sentence template semantic repair detection based on the text repair data, performing text semantic repair, and obtaining text sentence semantic repair data; Step S34: repair and improve the adversarial sample, specifically repairing and improving the adversarial sample through the numerical perturbation inverse repair, the text keyword normalization replacement and the text semantic repair, to obtain adversarial sample repair enhancement data.

6. A large-scale model data intelligent management method according to claim 5, characterized in that: In step S4, the data value assessment is used to dynamically quantify the contribution of data to model training. Specifically, based on the adversarial detection enhancement data and the adversarial sample repair enhancement data, a comprehensive normalization improved numerical value analysis method is used to perform data value assessment to obtain data value contribution assessment reference data, including the following steps: Step S41: numerical data value assessment, specifically, using a standard fitted linear regression model based on feature contribution inversion, and using the absolute value of the regression coefficient of each numerical data to the target variable as the prediction contribution score to perform numerical data value assessment and obtain numerical value assessment reference data; Step S42: evaluating the value of text data, specifically using a standard text information entropy evaluation method based on the TF-IDF matrix to evaluate the value of text data and obtain reference data for text value evaluation; Step S43: score normalization, specifically, by using a comprehensive normalization method to quantify and label the data value based on the numerical value evaluation reference data and the text value evaluation reference data, to obtain data value normalized reference data; The comprehensive normalization method includes quantitative normalization and label normalization, and the calculation formula is: ； Where S n is the quantitative normalized value assessment data, V n is the original value evaluation data extracted from the numerical value evaluation reference data and the text value evaluation reference data, min(·) is the minimum value obtaining function, V is the complete set of evaluation value data composed of the numerical value evaluation reference data and the text value evaluation reference data, max(·) is the maximum value obtaining function, and L n It is the label normalized evaluation data calculated based on the quantitative normalized value evaluation data; Step S44: Comprehensive evaluation of data value, specifically, performing comprehensive evaluation and normalization of data value through the numerical data value evaluation, the text data value evaluation and the score normalization to obtain reference data for data value contribution evaluation.

7. The method for intelligent management of large model data according to claim 6, characterized in that: In step S5, the large model data management is dynamically enhanced to combine adversarial detection, repair and data value evaluation to optimize the overall utilization of large model data resources. Specifically, the adversarial detection enhanced data and the adversarial sample repair enhanced data are combined to integrate the large model data source data to obtain fused enhanced structured data, and the data value is arranged according to the data value contribution evaluation reference data. By setting the dynamic threshold of value contribution, data is screened to obtain dynamically enhanced large model basic data.

8. A large model data intelligent management system, for implementing a large model data intelligent management method according to any one of claims 1 to 7, characterized in that: It includes data collection module, adversarial data detection module, adversarial sample repair module, data value assessment module and large model data management dynamic enhancement module.

9. The large-scale model data intelligent management system according to claim 8, characterized in that: The data collection module is used for data collection, obtains the original data of the structured large model through data collection, and sends the original data of the structured large model to the adversarial data detection module; The adversarial data detection module is used for adversarial data detection, obtains adversarial detection enhancement data through adversarial data detection, and sends the adversarial detection enhancement data to the adversarial sample repair module and the data value assessment module; The adversarial sample repair module is used to repair the adversarial sample, obtain adversarial sample repair enhancement data through adversarial sample repair, and send the adversarial sample repair enhancement data to the data value assessment module; The data value assessment module is used for data value assessment, obtains data value contribution assessment reference data through data value assessment, and sends the data value contribution assessment reference data to the large model data management dynamic enhancement module; The large model data management dynamic enhancement module is used for dynamic enhancement of large model data management, and obtains dynamically enhanced large model basic data through dynamic enhancement of large model data management.

Images