CN117521116B - Large language model privacy information protection method - Google Patents
Large language model privacy information protection method Download PDFInfo
- Publication number
- CN117521116B CN117521116B CN202410013413.7A CN202410013413A CN117521116B CN 117521116 B CN117521116 B CN 117521116B CN 202410013413 A CN202410013413 A CN 202410013413A CN 117521116 B CN117521116 B CN 117521116B
- Authority
- CN
- China
- Prior art keywords
- language model
- different
- word segmentation
- large language
- text sequence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 48
- 230000011218 segmentation Effects 0.000 claims abstract description 92
- 238000011156 evaluation Methods 0.000 claims abstract description 37
- 230000006870 function Effects 0.000 claims abstract description 23
- 238000000586 desensitisation Methods 0.000 claims abstract description 20
- 238000006467 substitution reaction Methods 0.000 claims abstract description 14
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 239000013598 vector Substances 0.000 claims description 32
- 230000008569 process Effects 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 8
- 238000012545 processing Methods 0.000 abstract description 10
- 239000012634 fragment Substances 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/602—Providing cryptographic facilities or services
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/237—Lexical tools
- G06F40/247—Thesauruses; Synonyms
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Bioethics (AREA)
- Data Mining & Analysis (AREA)
- Computer Hardware Design (AREA)
- Computer Security & Cryptography (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Databases & Information Systems (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- Probability & Statistics with Applications (AREA)
- Medical Informatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Machine Translation (AREA)
Abstract
The invention relates to the technical field of data processing, and provides a large language model privacy information protection method, which comprises the following steps: acquiring relevant parameters of a large language model; preprocessing the obtained large language model related parameters to obtain word segmentation data with different dimensions, calculating similar semantic sets of the word segmentation data with different dimensions according to the large language model related parameters, and calculating important association evaluation coefficients of word segmentation with different dimensions; acquiring synonym forest codes by using related parameters of a large language model, calculating associated similarity indexes of different dimension word segments at different moments, calculating a clustering evaluation function according to the associated similarity indexes of the different dimension word segments at different moments, and acquiring a high-frequency privacy information cluster; and acquiring a desensitization substitution text sequence according to the related parameters of the large language model by the high-frequency privacy information cluster, and protecting the privacy information of the large language model according to the desensitization substitution text sequence. The invention improves the reliability of protecting the privacy information of the large language model.
Description
Technical Field
The invention relates to the technical field of data processing, in particular to a large language model privacy information protection method.
Background
The large language model (Large Language Models, LLM) is a deep learning large model for realizing large-scale human natural language understanding through an artificial intelligence algorithm, thereby completing various tasks such as language reasoning, text generation, man-machine interaction and the like.
The large language model mainly represented by the dialogue generation and training model ChatGPT has various excellent capabilities such as high-quality dialogue, complex reasoning, cross-domain generalization and the like, and the large language model solves various open tasks through the excellent model performance assistance and is widely focused by different users in various different industrial fields.
Because the large language model parameter is very large in scale, the time cost and the time cost are large in the process of retraining the large language model, the general large language model widely applied in the industry at present mainly comes from a pre-trained general model disclosed on a network. The pre-training general model disclosed by the invention is quite similar to the text data processing process, and has excellent generalization, but the quite similar data processing flow also hides a larger risk of user privacy data disclosure.
Disclosure of Invention
The invention provides a large language model privacy information protection method, which solves the problem that the traditional clustering algorithm cannot accurately acquire privacy information, and adopts the following technical scheme:
the invention provides a large language model privacy information protection method, which comprises the following steps:
acquiring relevant parameters of a large language model;
Preprocessing the obtained related parameters of the large language model to obtain word segmentation data with different dimensions, calculating similar semantic sets of the word segmentation data with different dimensions according to the related parameters of the large language model, and calculating important association evaluation coefficients of the word segmentation with different dimensions according to the similar semantic sets;
Obtaining synonym Lin Bianma by using related parameters of a large language model, calculating associated similarity indexes of different dimension words at different moments according to important association coefficients of the different dimension words and synonym forest codes, calculating a clustering evaluation function according to the associated similarity indexes of the different dimension words at different moments, and obtaining a high-frequency privacy information cluster according to the clustering evaluation function;
And acquiring a desensitization substitution text sequence according to the related parameters of the large language model by the high-frequency privacy information cluster, and protecting the privacy information of the large language model according to the desensitization substitution text sequence.
Preferably, the large language model related parameters include: the user inputs a text sequence, a general text vector output by a large language model.
Preferably, the method for calculating the similar semantic collection of the word segmentation data with different dimensions according to the related parameters of the large language model comprises the following steps:
and marking the numerical value difference of the Chinese character codes of each dimension word segmentation data and other different dimension word segmentation as a first difference value, and forming a similar semantic set by all the different dimension word segmentation with the first difference value as a preset value.
Preferably, the mathematical expression for calculating the important association evaluation coefficients of the different dimension word segments according to the similar semantic sets is as follows:
in the/> Representing the adjustment of a preset constant,/>Represents the/>User input of the first/>, in a text sequence at a single instantWord frequency of individual dimension word segmentation data,/>Represents the/>User input text sequence at various moments/>Total number of all different word segmentation data of the same class semantic set of each dimension,/>Representing the relative distance between two different dimension word segmentation data,/>,/>Respectively show that at the/>User input of the first/>, in a text sequence at a single instantIndividual dimensions and/>Word segmentation data of individual dimensions,/>Represents the/>Middle/>, of a text sequence entered by a user at a single moment in timeImportant associated evaluation coefficients of individual dimension word segmentation.
Preferably, the mathematical expression for calculating the association similarity indexes of the different dimension words at different moments according to the important association coefficients of the different dimension words and the synonym forest codes is as follows:
in the/> Represents the/>User input of the first/>, in a text sequence at a single instantImportant association evaluation coefficient of individual dimension word segmentation,/>Represents the/>Dimension of user input text sequence data at various moments,/>Representing cosine similarity between two different dimension word-segmentation encoding vectors,/>Respectively represent the/>User input of the first text sequence at a timeIndividual dimension word segmentation and the/>Encoding vector of individual dimension word segmentation,/>An exponential function based on a natural constant is shown,Representing the height difference of synonym forest where two different dimension word segmentation coding vectors are located,/>, andRepresents the/>At the moment/>Associated similarity index for individual dimension word segments.
Preferably, the mathematical expression of the height difference calculation of the synonym forest where the two different dimension word segmentation coding vectors are located is:
in the/> ,/>Respectively represent the/>At the moment/>Individual dimension word segmentation and the/>The height of synonym forest where two different dimension word segmentation coding vectors are located,/>Representing a preset discriminating constant,/>,/>Respectively represent the/>At the moment/>Individual dimensions and/>A set formed by all different synonyms in a synonym forest where each dimension word is located,/>Represents the/>User input of the first/>, in a text sequence at a single instantIndividual dimension word segmentation and the/>Height difference of synonym forest among individual dimension word segments.
Preferably, the mathematical expression of the clustering evaluation function calculated according to the associated similarity indexes of the different dimension word segments at different moments is:
in the/> Represents the/>Dimension of text sequence data entered by a user at each instant,/>Representing the total number of categories of the text data sequence input by the user,/>Represents the/>At the moment/>Associated similarity index of individual dimension word segment,/>Represents the/>Class center dimension word segmentation/>, at each momentRelated similarity index,/>Represents the/>The text sequence data entered by the user at each instant clusters the evaluation function.
Preferably, the method for acquiring the high-frequency privacy information cluster according to the clustering evaluation function comprises the following steps:
And clustering the text sequence data to obtain a cluster when the minimum value of the evaluation function is clustered, and marking the cluster as a high-frequency privacy information cluster.
Preferably, the method for acquiring the desensitization substitution text sequence according to the related parameters of the high-frequency privacy information cluster to the large language model comprises the following steps:
and replacing the high-frequency privacy information cluster in the related parameters of the large language model in the Greek alphabet according to a preset desensitization step length, and recording the data obtained after replacement as a desensitization replacement text sequence.
Preferably, the method for protecting the privacy information of the large language model according to the desensitization substitution text sequence comprises the following steps:
and inputting the desensitization substitution text sequence as an AES encryption algorithm to obtain the encrypted large language model related parameters, so as to protect the privacy information of the large language model.
The beneficial effects of the invention are as follows: according to the invention, different similar semantic sets are constructed through the word coding differences of the word segmentation in different dimensions in the large language model, and important association evaluation coefficients are obtained by calculating different word segmentation data in the similar semantic sets, so that the privacy information in the text data input by the large language model user is primarily extracted and calculated. Furthermore, the method acquires the coding vectors of the different-dimension word fragments through the synonym forest, calculates the association similarity indexes of the different-dimension word fragments by combining the important association evaluation coefficients and the coding vectors, and effectively reflects the privacy information data in the text data input by the large language model user through the association similarity indexes. According to the method, the traditional clustering loss function is optimized by combining the privacy information data, so that the problem of inaccurate acquisition of the privacy information caused by direct clustering of text information input by a large language model user in a traditional clustering algorithm is effectively solved.
Drawings
In order to more clearly illustrate the embodiments of the invention or the technical solutions of the prior art, the drawings which are used in the description of the embodiments or the prior art will be briefly described, it being obvious that the drawings in the description below are only some embodiments of the invention, and that other drawings can be obtained according to these drawings without inventive faculty for a person skilled in the art.
FIG. 1 is a flowchart of a method for protecting privacy information of a large language model according to an embodiment of the present invention;
Fig. 2 is a schematic diagram of a correlation similarity index calculation flow.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Referring to fig. 1, a flowchart of a method for protecting privacy information of a large language model according to an embodiment of the invention is shown, the method includes the following steps:
Step S001, acquiring relevant parameters of the large language model.
It should be noted that, in general, the main structure of the general large language model is stacked layer by layer according to the network structure model of the Transformer, and the text sequence input by the user is processed by the large language model to generate the corresponding general text vector feature, and different language processing tasks are further completed according to the general text vector. Therefore, in the processing process of the large language model, if the user unintentionally leaks the cached general text vector output by the large language model in the process of using the large language model, or an attacker acquires the general text vector output by the large language model in various modes such as acquiring the general text vector output by the large language model by capturing data information of a victim server through an attack channel. An attacker can infer the original text data input by the user according to the universal text vector to obtain the user privacy key sensitive words.
In order to protect the privacy information of the large language model, related parameters of the large language model need to be acquired, wherein the related parameters comprise a text sequence input by a user and a general text vector output by the large language model.
Considering that the language types used by users in different regions are different, large language model parameters can have large difference in the processing process of different language types, so that the invention firstly processes the original text data input by the users and converts the non-Chinese original text data into Chinese text data. Meanwhile, in order to facilitate subsequent further processing of the original text data, the text sequence input by the user is subjected to word segmentation processing through the jieba word segmentation tool to obtain different word segmentation data, and in order to avoid influence of high-frequency stop words in the original text data on subsequent further analysis and calculation processes, the stop words in the original text data are removed through the word segmentation tool.
Step S002, preprocessing the obtained related parameters of the large language model to obtain word segmentation data with different dimensions, calculating similar semantic sets of the word segmentation data with different dimensions according to the related parameters of the large language model, and calculating important association evaluation coefficients of the word segmentation with different dimensions according to the similar semantic sets.
It should be noted that, in general, the large language model directly analyzes and processes the text sequence data input by the user to obtain a corresponding general text vector. However, when text sequence data input by a user is directly processed through a large language model, privacy information in original text sequence data of the user is exposed to a large extent, so that analysis of the text sequence data input by the user is required in order to avoid the defect of exposure of privacy information when an attacker attacks the large language model.
Assume time of dayThe dimension of text sequence data input by a user is/>The user-entered text sequence data at this point may be noted as/>Wherein/>,/>,/>Respectively represent time/>At user input of text sequence data/>Dimension, th/>Sum/>Dimension. Each different dimension of the user-entered text sequence data represents a different word segmentation data.
For user input of text sequence dataWord segmentation data for each dimension of (1)Calculating the Chinese character encoding numerical value difference between the word segmentation data of each different dimension and the word segmentation data of other dimensions, and if the two different Chinese characters are identical in representation in a computer, the Chinese character encoding numerical values between the two characters are identical, so that the Chinese character encoding numerical value difference of the word segmentation data of each different dimension is/>The invention uses GBK Chinese character codes to calculate, in specific application, the implementer can set according to specific conditions to set time/>, and the word segmentation data is used as the similar semantic set of corresponding dimension word segmentation dataAt user input text sequence number/>The homogeneous semantic collections of individual dimensions are denoted/>。
In the/>Representing the adjustment of a preset constant,/>Represents the/>User input of the first/>, in a text sequence at a single instantWord frequency of individual dimension word segmentation data,/>Represents the/>User input text sequence at various moments/>Total number of all different word segmentation data of the same class semantic set of each dimension,/>Representing the relative distance between two different dimension word segmentation data,/>,/>Respectively show that at the/>User input of the first/>, in a text sequence at a single instantIndividual dimensions and/>Word segmentation data of individual dimensions,/>Represents the/>Middle/>, of a text sequence entered by a user at a single moment in timeImportant associated evaluation coefficients of individual dimension word segmentation.
The time can be set by the methodCalculating important associated evaluation coefficients of word segmentation data with different dimensions in a text sequence input by a user, wherein a preset constant/>Take the empirical value as/>Approximate distance/>, between two different dimension word segmentation dataThe calculation method is the total number of the word segmentation data of all other dimensions between two different word segmentation data. If/>The/>, in the text sequence entered by the user at each instantThe higher the frequency of occurrence of individual segmentation, the more frequently the word is simultaneously with the/>The smaller the interval between other words with the same word segmentation, the description time/>Input text sequence data by userThe more important the individual dimension word segmentation data is, the more likely it is that the key privacy information in the text data is input by the user, then the/>The larger the value of the important association evaluation coefficient obtained by calculating the individual dimension word segmentation data.
And S003, acquiring synonym Lin Bianma by using related parameters of the large language model, calculating associated similarity indexes of the different-dimension words at different moments according to important association coefficients of the different-dimension words and synonym forest codes, calculating a clustering evaluation function according to the associated similarity indexes of the different-dimension words at different moments, and acquiring a high-frequency privacy information cluster according to the clustering evaluation function.
When calculating the text sequence input by the user, the first text code is obtained and input in the text sequenceSimilar semantic sets of individual dimension word segments, however, different synonym transformation descriptions may exist for the same dimension word segment in text sequence data input by a user, and similar synonyms cannot be effectively judged through strict Chinese character coding comparison calculation, so that calculation analysis is required for synonym similarity among different dimension word segments in text data input by the user.
Because different words have different associated interpretations, the method aims at carrying out associated analysis on synonym information among different dimension word segments in text data input by a user. As shown in fig. 2, the method encodes all the different dimension words in the text data sequence input by the user according to the "synonym forest" to obtain corresponding synonyms Lin Bianma, the "synonym forest" and its specific encoding process are known techniques, and are not described in detail herein, and the moment is assumedFirst/>, in text data entered by a user at a locationThe code vector of each dimension word is recorded as/>。
In the/>Represents the/>User input of the first/>, in a text sequence at a single instantImportant association evaluation coefficient of individual dimension word segmentation,/>Represents the/>Dimension of user input text sequence data at various moments,/>Representing cosine similarity between two different dimension word-segmentation encoding vectors,/>Respectively represent the/>User input of the first text sequence at a timeIndividual dimension word segmentation and the/>Encoding vector of individual dimension word segmentation,/>An exponential function based on a natural constant is shown,Representing the height difference of synonym forest where two different dimension word segmentation coding vectors are located,/>, andRepresents the/>At the moment/>Associated similarity index for individual dimension word segments.
The synonymous association index of the text sequences input by the user at different moments can be calculated through the method, if the moments areCalculation of the place to obtain the/>The larger the important association evaluation coefficient of each dimension word is, the higher the cosine similarity between two different dimension word coding vectors is and the smaller the height difference value of the corresponding different dimension word in the synonym forest is, the description is in the/>User input of the first/>, in a text sequence at a single instantThe higher the possibility that the individual dimension word is user privacy information, the more/>, the more the user privacy information is calculated at the momentAt the moment/>Individual dimension word association similarity index/>The larger the value of (c) will be relatively.
In the/>,/>Respectively represent the/>At the moment/>Individual dimension word segmentation and the/>The height of synonym forest where two different dimension word segmentation coding vectors are located,/>Representing a preset discriminating constant,/>,/>Respectively represent the/>At the moment/>Individual dimensions and/>A set formed by all different synonyms in a synonym forest where each dimension word is located,/>Represents the/>User input of the first/>, in a text sequence at a single instantIndividual dimension word segmentation and the/>Height difference of synonym forest among individual dimension word segments.
The height difference of the synonym forest where the two different-dimension word segmentation coding vectors are located can be calculated through the method, and the fact that the synonym forest where the two different-dimension word segmentation coding vectors are located is large is assumed that the difference of the synonym forest where the two different-dimension word segmentation coding vectors are located is large, at the moment, the intersection between sets formed by all different synonyms in the corresponding synonym forest is empty, the fact that the association coincidence degree between the two different-dimension word segmentation coding vectors is small is indicated, two different semantic sets are represented, and at the moment, the preset difference constant is larger experience value isThe numerical value of the height difference of the synonym forest calculated in the same synonym forest should be larger than that of the height difference of the synonym forest calculated in the same synonym forest, and in specific application, an implementer can set according to specific situations.
It should be noted that, for the text sequence input by the user at different moments, the similar words appearing repeatedly and frequently are very likely to represent the key privacy information of the user, so the invention makesTake the experience value/>In a specific application, the implementer may make settings according to the specific circumstances. Clustering the text sequences input by the user by using a K-Means algorithm to obtain the most important/>, of the privacy information of the data input by the userThe categories.
In the/>Represents the/>Dimension of text sequence data entered by a user at each instant,/>Representing the total number of categories of the text data sequence input by the user,/>Represents the/>At the moment/>Associated similarity index of individual dimension word segment,/>Represents the/>Class center dimension word segmentation/>, at each momentRelated similarity index,/>Represents the/>The text sequence data entered by the user at each instant clusters the evaluation function.
The time can be calculated by the above methodThe clustering evaluation function of text sequence data is input by a user, and the method is distributed in/>, according to Gaussian distributionRandom selection/>, among individual different dimension segmentationsAnd taking the word segmentation data with different dimensions and the class center word segmentation as initial clustering centers, and assuming that the smaller the difference of the association similarity indexes of the word segmentation data with different dimensions and the class center word segmentation is, the current dimension word segmentation and the class center word segmentation are both in the same semantic representation range. Thus, when calculated/>Text sequence data clustering evaluation function/>, input by user at each momentWhen the numerical value is minimum, the convergence of the current clustering function is illustrated, and the text sequence data input by the user is clustered and divided to obtain/>Different high frequency privacy information clusters.
And S004, acquiring a desensitization substitution text sequence for the related parameters of the large language model according to the high-frequency privacy information cluster, and protecting the privacy information of the large language model according to the desensitization substitution text sequence.
It should be noted that, for text sequence data input by the user at different time points, it is possible to obtainDifferent high-frequency privacy information clusters which contain privacy information in text sequence data input by a user, so that the invention combines the alphabetic permutation pairs/>The different high-frequency privacy information clusters are subjected to desensitization protection.
The invention takes the preset desensitization step lengthFor/>Pairs/>, in 24 different greek alphabetsThe private data in different high-frequency private information clusters are mapped and replaced to obtain desensitized replacement text sequence data input by users at different moments, the desensitization step length is preset, the desensitization replacement text sequence data input by the users at different moments are processed through an AES encryption algorithm, and the encrypted text sequence obtained after encryption is sent to a large language model server for further processing. By desensitizing and encrypting the user transmitted data at different moments, the security of the privacy data of the large language model in the text sequence data input by the user is ensured, and the risk of intercepting the text sequence data input by the user in the channel attack process by an attacker is avoided.
The above embodiments are only for illustrating the technical solution of the present application, and not for limiting the same; although the application has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the application and are intended to be included within the scope of the application.
Claims (9)
1. A method for protecting privacy information of a large language model, comprising the steps of:
acquiring relevant parameters of a large language model;
Preprocessing the obtained related parameters of the large language model to obtain word segmentation data with different dimensions, calculating similar semantic sets of the word segmentation data with different dimensions according to the related parameters of the large language model, and calculating important association evaluation coefficients of the word segmentation with different dimensions according to the similar semantic sets;
Obtaining synonym Lin Bianma by using related parameters of a large language model, calculating associated similarity indexes of different dimension words at different moments according to important associated evaluation coefficients of the different dimension words and synonym forest codes, calculating a cluster evaluation function according to the associated similarity indexes of the different dimension words at different moments, and obtaining a high-frequency privacy information cluster according to the cluster evaluation function;
Acquiring a desensitization substitution text sequence for related parameters of the large language model according to the high-frequency privacy information cluster, and protecting the privacy information of the large language model according to the desensitization substitution text sequence;
The large language model related parameters include: the user inputs a text sequence, a general text vector output by a large language model.
2. The method for protecting privacy information of large language model according to claim 1, wherein the method for calculating the similar semantic collection of word segmentation data with different dimensions according to the related parameters of the large language model is as follows:
and marking the numerical value difference of the Chinese character codes of each dimension word segmentation data and other different dimension word segmentation as a first difference value, and forming a similar semantic set by all the different dimension word segmentation with the first difference value as a preset value.
3. The method for protecting privacy information of large language model according to claim 2, wherein the mathematical expression for calculating the important associated evaluation coefficients of different dimension word segments according to the similar semantic collection is:
in the/> Indicating that the adjustment of the preset constant is performed,Represents the/>User input of the first/>, in a text sequence at a single instantWord frequency of individual dimension word segmentation data,/>Represents the/>User input text sequence at various moments/>The total number of all different word segmentation data of the same class semantic set of each dimension,Representing the relative distance between two different dimension word segmentation data,/>,/>Respectively show that at the/>User input of the first/>, in a text sequence at a single instantIndividual dimensions and/>Word segmentation data of individual dimensions,/>Represents the/>Middle/>, of a text sequence entered by a user at a single moment in timeImportant associated evaluation coefficients of individual dimension word segmentation.
4. The method for protecting privacy information of large language model according to claim 3, wherein the mathematical expression of the association similarity index of the different dimension word segments at different time points according to the important association evaluation coefficient of the different dimension word segments and the synonym word forest code is calculated as follows:
in the/> Represents the/>User input of the first/>, in a text sequence at a single instantImportant associated evaluation coefficients of individual dimension word segmentation,Represents the/>Dimension of user input text sequence data at various moments,/>Representing cosine similarity between two different dimension word-segmentation encoding vectors,/>Respectively represent the/>User input of the first/>, in a text sequence at a single instantIndividual dimension word segmentation and the/>Encoding vector of individual dimension word segmentation,/>Represents an exponential function based on natural constants,/>Representing the height difference of synonym forest where two different dimension word segmentation coding vectors are located,/>, andRepresents the/>At the moment/>Associated similarity index for individual dimension word segments.
5. The method for protecting private information of large language model according to claim 4, wherein the mathematical expression of the calculation of the height difference of the synonym forest where the two different dimension word segmentation coding vectors are located is:
In the method, in the process of the invention, ,/>Respectively represent the/>At the moment/>Individual dimension word segmentation and the/>The height of synonym forest where two different dimension word segmentation coding vectors are located,/>Representing a preset discriminating constant,/>,/>Respectively represent the/>At the moment/>Individual dimensions and/>A set of all different synonyms in the synonym forest where each dimension word is located,Represents the/>User input of the first/>, in a text sequence at a single instantIndividual dimension word segmentation and the/>Height difference of synonym forest among individual dimension word segments.
6. The method for protecting privacy information of large language model according to claim 4, wherein the mathematical expression of the clustering evaluation function is calculated according to the associated similarity indexes of different dimension word segments at different moments:
in the/> Represents the/>Dimension of text sequence data entered by a user at each instant,/>Representing the total number of categories of the text data sequence input by the user,/>Represents the/>At the moment/>Associated similarity index of individual dimension word segment,/>Represents the/>Class center dimension word segmentation/>, at each momentRelated similarity index,/>Represents the/>The text sequence data entered by the user at each instant clusters the evaluation function.
7. The method for protecting private information of large language model according to claim 6, wherein the method for obtaining high-frequency private information clusters according to the cluster evaluation function is as follows:
And clustering the text sequence data to obtain a cluster when the minimum value of the evaluation function is clustered, and marking the cluster as a high-frequency privacy information cluster.
8. The method for protecting privacy information of large language model according to claim 7, wherein the method for obtaining desensitized substitution text sequence for related parameters of large language model according to high frequency privacy information cluster comprises:
and replacing the high-frequency privacy information cluster in the related parameters of the large language model in the Greek alphabet according to a preset desensitization step length, and recording the data obtained after replacement as a desensitization replacement text sequence.
9. The method for protecting private information of a large language model according to claim 1, wherein the method for protecting private information of a large language model according to the desensitized alternative text sequence comprises the steps of:
and inputting the desensitization substitution text sequence as an AES encryption algorithm to obtain the encrypted large language model related parameters, so as to protect the privacy information of the large language model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410013413.7A CN117521116B (en) | 2024-01-04 | 2024-01-04 | Large language model privacy information protection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202410013413.7A CN117521116B (en) | 2024-01-04 | 2024-01-04 | Large language model privacy information protection method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117521116A CN117521116A (en) | 2024-02-06 |
CN117521116B true CN117521116B (en) | 2024-04-19 |
Family
ID=89745992
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202410013413.7A Active CN117521116B (en) | 2024-01-04 | 2024-01-04 | Large language model privacy information protection method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117521116B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117808643B (en) * | 2024-02-29 | 2024-05-28 | 四川师范大学 | Teaching management system based on Chinese language |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113468885A (en) * | 2021-07-13 | 2021-10-01 | 安徽大学绿色产业创新研究院 | Chinese trademark similarity calculation method |
CN116975927A (en) * | 2023-08-17 | 2023-10-31 | 南开大学 | LLM language user privacy information protection method based on natural language prompt |
CN117313138A (en) * | 2023-08-30 | 2023-12-29 | 西安电子科技大学 | Social network privacy sensing system and method based on NLP |
-
2024
- 2024-01-04 CN CN202410013413.7A patent/CN117521116B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113468885A (en) * | 2021-07-13 | 2021-10-01 | 安徽大学绿色产业创新研究院 | Chinese trademark similarity calculation method |
CN116975927A (en) * | 2023-08-17 | 2023-10-31 | 南开大学 | LLM language user privacy information protection method based on natural language prompt |
CN117313138A (en) * | 2023-08-30 | 2023-12-29 | 西安电子科技大学 | Social network privacy sensing system and method based on NLP |
Non-Patent Citations (4)
Title |
---|
基于语义关联与模糊聚类的共词分析方法;陆泉 等;情报学报;20221031;第41卷(第10期);第1003-1014页 * |
改进TF-IDF结合余弦定理计算中文语句相似度;张俊飞;;现代计算机(专业版);20171115(第32期);第20-23、27页 * |
融合词语共现距离和类别信息的短文本特征提取方法;马慧芳 等;计算机工程与科学;20180915(第09期);第1689-1695页 * |
语义相似和多维加权的联合敏感属性隐私保护;徐龙琴 等;计算机应用;20110401(第04期);第999-1002页 * |
Also Published As
Publication number | Publication date |
---|---|
CN117521116A (en) | 2024-02-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN117521116B (en) | Large language model privacy information protection method | |
CN109308494B (en) | LSTM model and network attack identification method and system based on LSTM model | |
Qin et al. | A network security entity recognition method based on feature template and CNN-BiLSTM-CRF | |
CN111241291A (en) | Method and device for generating countermeasure sample by utilizing countermeasure generation network | |
CN110019758B (en) | Core element extraction method and device and electronic equipment | |
CN110363001B (en) | Application layer malicious request detection method based on Transformer model | |
CN113315789B (en) | Web attack detection method and system based on multi-level combined network | |
CN112651025A (en) | Webshell detection method based on character-level embedded code | |
CN115146068B (en) | Method, device, equipment and storage medium for extracting relation triples | |
CN113094478A (en) | Expression reply method, device, equipment and storage medium | |
CN110674370A (en) | Domain name identification method and device, storage medium and electronic equipment | |
CN114547670A (en) | Sensitive text desensitization method using differential privacy word embedding disturbance | |
CN113742763A (en) | Confusion encryption method and system based on government affair sensitive data | |
CN111737688B (en) | Attack defense system based on user portrait | |
CN117332411A (en) | Abnormal login detection method based on transducer model | |
CN112711648A (en) | Database character string ciphertext storage method, electronic device and medium | |
CN115268799B (en) | Storage method and device based on cloud service | |
CN117271759A (en) | Text abstract generation model training method, text abstract generation method and device | |
CN112507388B (en) | Word2vec model training method, device and system based on privacy protection | |
CN113822018A (en) | Entity relation joint extraction method | |
CN114338058A (en) | Information processing method, device and storage medium | |
CN112182575A (en) | Attack data set malicious segment marking method and system based on LSTM | |
CN116756296B (en) | Consultation information management method and system based on privacy protection | |
CN116611037B (en) | Deep neural network black box watermarking method, device and terminal | |
CN115600580B (en) | Text matching method, device, equipment and storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |