Privacy protection data processing system and method for large language model
By extracting statistical features from data in large-scale model applications and encrypting them, the contradiction between privacy protection and data availability is resolved, enabling secure and effective analysis in highly sensitive data scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-03-31
AI Technical Summary
Existing privacy protection methods in large-scale model applications suffer from a trade-off between sacrificing data privacy, data availability, and communication overhead and performance, and cannot meet the needs of highly sensitive data scenarios.
The architecture consists of a user environment layer, a local MCP Server layer, and a large model service layer. By extracting statistical features from the data and encrypting them, the large model only analyzes encrypted data, avoiding direct access to the original data.
It achieves a balance between data privacy protection and model utilization, enhances privacy protection strength, data availability and system performance, and meets the security requirements of highly sensitive industries.
Smart Images
Figure CN121765760A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and data security, and specifically relates to a privacy-preserving data processing system and method for large language models. Background Technology
[0002] With the development of digital technology, large language model (hereinafter referred to as large model) technology has been deeply applied in highly sensitive industries such as finance, healthcare, and government affairs due to its powerful analysis and processing capabilities. In the financial sector, large models are used in core business processes such as risk assessment, credit scoring, and investment decision-making, helping financial institutions improve decision-making efficiency and accuracy. In the healthcare sector, large models assist in disease diagnosis, drug development, and patient management, improving the quality and efficiency of medical services. In the government sector, large models are used for policy analysis and public service optimization, helping governments improve governance capabilities and service levels. However, the data in these industries often contains a large amount of personal privacy and trade secrets, and data privacy protection has become a bottleneck restricting the application of large model technology.
[0003] Currently, in the application of large model technology, large model services usually need to access raw data for training and inference, which leads to the risk of sensitive data leakage. At the same time, the application of privacy protection technology often reduces data availability and affects the analysis effect of large models. Therefore, how to protect data privacy while maintaining the statistical characteristics of data and supporting the effective analysis of large models has become an urgent technical problem to be solved.
[0004] Currently, solutions to privacy issues in large-scale model data analysis mainly include direct cloud service calls, traditional data anonymization techniques, and federated learning techniques. Direct cloud service calls involve users directly accessing cloud services provided by third-party large-scale model vendors via API interfaces, with the raw data transmitted in plaintext to the vendor's cloud server for processing. Traditional data anonymization techniques de-identify sensitive information before data upload, such as substitution anonymization and generalization anonymization. Federated learning techniques protect the raw data by training it locally and sharing model parameters; each participant trains the model on local data, only transmitting the model parameters to a central server for aggregation.
[0005] However, the aforementioned traditional privacy protection methods all rely on protecting the raw data itself, which presents a fundamental contradiction between privacy protection and large-scale model utilization: sacrificing data privacy and availability, and facing communication overhead and performance issues. Specifically, direct access to cloud services carries inherent data leakage risks and significant compliance risks; traditional data anonymization techniques are prone to severe data value loss, have limited applicability, and offer insufficient privacy protection; federated learning techniques face problems such as huge communication overhead, degraded model performance, and high implementation complexity, and are only suitable for multi-party data interaction training, with poor performance from a single data source. These contradictions limit the application of large-scale model technology in high-value and sensitive data processing scenarios such as finance and healthcare, failing to simultaneously meet the dual requirements of strong privacy protection and high data availability, unable to fully leverage the analytical capabilities of large models while ensuring data security, and exhibiting excessive technical complexity, making it difficult to meet increasingly stringent data protection regulations. Therefore, there is an urgent need to develop a new privacy-preserving data processing technology solution. Summary of the Invention
[0006] To address the technical problems existing in the prior art, this invention provides a privacy-preserving data processing system and method for large language models, in order to solve the technical problems of traditional privacy protection methods sacrificing data privacy and availability, as well as facing communication overhead and performance issues, in the balance between privacy protection and large model utilization.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a privacy-preserving data processing system for large language models, comprising a user environment layer, a local MCP Server layer, and a large model service layer; The user environment layer is used to store raw data locally and receive requests from business applications. The local MCP Server layer is used to extract the data to be analyzed from the raw data according to the business application request; encrypt the data to be analyzed according to the pre-extracted statistical features to obtain encrypted data; and perform preset data operation processing on the data to be analyzed based on the processing instructions issued by the large language model. The large model service layer is used to analyze encrypted data using large language models and issue processing instructions to the local MCP Server layer.
[0008] Furthermore, the local MCP Server layer includes an instruction parsing module, MCP privacy protection tools, and a local data processing engine; The instruction parsing module is used to receive and parse business application requests to obtain the parsed business application requests; it is also used to receive and parse the processing instructions issued by the large language model and pass the parsed processing instructions to the MCP privacy protection tool. The MCP privacy protection tool is used to extract data to be analyzed from the raw data based on the parsed business application requests, and to extract statistical features from the data to be analyzed; based on the extracted statistical features, the data to be analyzed is encrypted to obtain encrypted data; it is also used to identify parameters of the parsed processing instructions to obtain instruction parameter identification results. The local data processing engine is used to identify the results based on the command parameters and perform preset data operations on the data to be analyzed.
[0009] Furthermore, the MCP privacy protection tools include statistical analysis components, reversible encryption components, and parameter identification components; The statistical analysis component is used to extract the data to be analyzed from the raw data based on the parsed business application requests, extract the preset statistical features of the data to be analyzed, and obtain the extracted statistical features. A reversible encryption component is used to reversibly encrypt the data to be analyzed based on extracted statistical features to obtain encrypted data. The parameter recognition component is used to identify the parameters in the parsed processing instructions and obtain the parameter recognition results; analyze the semantic context of the parameters in the parsed processing instructions to obtain the parameter data type; and generate the instruction parameter recognition results based on the parameter data type and the parameter recognition results.
[0010] Furthermore, the preset statistical characteristics of the data to be analyzed include numerical characteristics, categorical characteristics, and correlation indicators between characteristics; Numerical characteristics include mean, median, standard deviation, quartiles, and distribution type; Categorical features include frequency distribution, mode, and cardinality; The correlation indicators of features include the correlation coefficient and mutual information between features.
[0011] Furthermore, based on the extracted statistical features, the data to be analyzed is reversibly encrypted to obtain encrypted data, as follows: For numerical data to be analyzed, a linear transformation is performed on the data based on the extracted statistical features, and differential privacy noise is added; the differential privacy noise is calculated based on the extracted statistical features. For categorical data to be analyzed, k-anonymization is performed on the data based on the extracted statistical features. For text-based data to be analyzed, calibration noise is added to the word embedding space of the data; the calibration noise is calculated based on the extracted statistical features.
[0012] Furthermore, the process of generating instruction parameter recognition results based on parameter data type and parameter recognition results is as follows: When the parameter data type is a real business value, the parameter recognition result will be directly output as the instruction parameter recognition result; When the parameter data type is a statistical inference value, the parameter identification result is mapped to the true value range using a pre-calculated statistical mapping table, and the instruction parameter identification result is output.
[0013] Furthermore, the preset data processing operations include data filtering, aggregation calculation, correlation analysis, and feature engineering.
[0014] Furthermore, the instruction parsing module supports parsing natural language instructions and structured query languages; the local data processing engine supports processing structured data, semi-structured data, and unstructured data.
[0015] Furthermore, the raw data can be financial transaction data, government data, or medical data.
[0016] The present invention also provides a privacy-preserving data processing method for large language models, utilizing the aforementioned privacy-preserving data processing system for large language models; Store the raw data locally and receive requests from business applications; Based on business application requests, extract the data to be analyzed from the raw data; encrypt the data to be analyzed based on the pre-extracted statistical features to obtain encrypted data; and use it to perform preset data operation processing on the data to be analyzed based on the processing instructions issued by the large language model. The encrypted data is analyzed using a large language model, and processing instructions are issued.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The privacy-preserving data processing system for large language models provided by this invention encrypts the data to be analyzed based on pre-extracted statistical features, and then analyzes the encrypted data by the large model service layer, achieving a balance between data privacy protection and model utilization, and providing technical support for large model applications in privacy-sensitive scenarios. Specifically, the system utilizes the user environment layer for local storage of raw data and access for business application requests, ensuring the physical isolation of the raw data; at the local MCPServer layer, the data to be analyzed is encrypted according to the pre-extracted statistical features, so that the encrypted data retains the statistical characteristics of the data, maintaining data availability while protecting privacy, thus achieving a balance between data privacy protection and data availability; the encrypted data is analyzed in the large model service layer and processing instructions are returned, and then processed by the local MCP... The server layer performs data manipulation operations on the raw data, while the large model service layer only receives privacy-protected data and cannot access the raw data, eliminating the risk of raw data leakage and ensuring an effective balance between data privacy protection and large model analysis capabilities. In this invention, the raw data is encrypted based on statistical features, achieving a balance between data privacy protection and data availability. The large model service layer can only access statistical features and encrypted data, and the raw data is always processed locally. This significantly improves key indicators such as privacy protection strength, data availability, and system performance, completely eliminating the risk of data leakage and meeting the stringent data security requirements of highly sensitive industries such as finance, government, and healthcare. Simultaneously, it effectively maintains the accuracy and reliability of business decisions, reduces the amount of data transmitted over the network, improves processing efficiency, and supports the needs of real-time business scenarios.
[0018] The privacy-preserving data processing method for large language models provided by this invention possesses all the advantages of the aforementioned privacy-preserving data processing system for large language models. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a structural block diagram of the privacy-preserving data processing system for large language models provided in Example 1; Figure 2 This is a schematic diagram illustrating the working principle of the MCP privacy protection tool in Example 1; Figure 3 This is a flowchart illustrating the workflow of the privacy-preserving data processing system for large language models provided in Example 1 in a financial risk control scenario. Figure 4 This is a flowchart of a privacy-preserving data processing method for a large language model provided in Example 2. Detailed Implementation
[0021] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0022] Before describing the specific embodiments of this application, some of the technical terms involved in the embodiments of this application are explained as follows: Large Language Model (LLM) is a deep learning model trained on massive amounts of text data. Its core objective is to generate and understand natural language text by learning the patterns and structures of language.
[0023] MCP (Model Context Protocol) is an open-source communication standard designed to standardize the interaction interface between large models and external tools and data sources. By defining a unified protocol specification, it enables bidirectional communication and data integration between large models and heterogeneous resources (such as databases, APIs, and local software).
[0024] Reversible encryption technology is an encryption method that allows ciphertext to be decrypted to recover the original data; it uses symmetric or asymmetric encryption algorithms to encrypt and decrypt data, ensuring the security of data during transmission and storage.
[0025] Homomorphic encryption is an encryption method that allows specific operations (such as addition and multiplication) to be performed on ciphertext without decryption. It maintains the homomorphism between ciphertext and plaintext in terms of operations, that is, the result of performing the operation on the ciphertext is the same as the result of performing the same operation on the plaintext and then encrypting it.
[0026] k-anonymization is a data anonymization technique that aims to reduce the precision of published data through generalization and concealment techniques, ensuring that there are at least k records for the same quasi-identifier, thereby reducing the risk of privacy leaks caused by link attacks.
[0027] This invention provides a privacy-preserving data processing system for large language models, comprising a user environment layer, a local MCP Server layer, and a large model service layer. The user environment layer stores raw data locally and receives requests from business applications. The local MCP Server layer extracts data to be analyzed from the raw data according to the business application requests; encrypts the data to be analyzed based on pre-extracted statistical features to obtain encrypted data; and performs preset data operations on the data to be analyzed based on processing instructions issued by the large language model. The large model service layer analyzes the encrypted data using the large language model and issues processing instructions to the local MCP Server layer.
[0028] In the above implementation, the local MCP Server layer encrypts the data to be analyzed based on the pre-extracted statistical features, enabling a privacy-preserving transformation before the data is sent to the large model service layer. The large language model in the large model service layer performs data analysis without touching the original data. This approach protects data privacy while maintaining data statistical characteristics, supports effective large model analysis while ensuring data security, and is easy to implement while meeting compliance requirements. It achieves a balance between data privacy protection and large model utilization, providing technical support for large model applications in privacy-sensitive scenarios.
[0029] The following specific embodiments further illustrate the privacy-preserving data processing system for large language models provided by the present invention: Example 1 As attached Figure 1 As shown, this embodiment 1 provides a privacy-preserving data processing system for large language models, including a user environment layer, a local MCP Server layer, and a large model service layer; wherein, the user environment layer, the local MCP Server layer, and the large model service layer are all deployed locally, that is, the system described in this embodiment 1 adopts a fully localized deployment architecture.
[0030] The user environment layer is used to store raw data locally and receive requests from business applications to ensure the physical isolation of the raw data. Raw data includes, for example, financial transaction data, government data, or medical data. The local MCPServer layer is used to extract the data to be analyzed from the raw data according to the requests from business applications; encrypt the data to be analyzed according to the pre-extracted statistical features to obtain encrypted data; and perform preset data operation processing on the data to be analyzed based on the processing instructions issued by the large language model. The large model service layer is used to analyze the encrypted data using the large language model and issue processing instructions to the local MCPServer layer.
[0031] In this embodiment 1, the user environment layer includes a raw data storage module and a business application interface; the raw data storage module is deployed locally and is used to store, for example, financial data, government data and medical data; the business application interface is used to receive business application requests.
[0032] In this embodiment 1, the local MCP Server layer includes an instruction parsing module, an MCP privacy protection tool, and a local data processing engine. The instruction parsing module is used to receive and parse business application requests to obtain the parsed business application requests; it is also used to receive and parse processing instructions issued by the large language model and transmit the parsed processing instructions to the MCP privacy protection tool. The MCP privacy protection tool is used to extract the data to be analyzed from the original data according to the parsed business application requests, and to extract statistical features from the data to be analyzed; it encrypts the data to be analyzed based on the extracted statistical features to obtain encrypted data; it is also used to perform parameter recognition on the parsed processing instructions to obtain instruction parameter recognition results. The local data processing engine is used to perform preset data operation processing on the data to be analyzed according to the instruction parameter recognition results; wherein, the preset data operation processing includes data filtering operation, aggregation calculation operation, correlation analysis operation, and feature engineering operation.
[0033] In detail, within the MCP Server layer, the MCP privacy protection tool achieves a balance between data privacy protection and data availability, working collaboratively with the instruction parsing module and the local data processing engine to complete privacy-protected data processing operations. Preferably, the MCP privacy protection tool employs a distributed deployment architecture to support large-scale data processing scenarios. The nodes in this distributed deployment architecture are equipped with secure communication mechanisms and data consistency guarantee strategies. (See attached...) Figure 2 As shown, the MCP privacy protection tool includes a statistical analysis component, a reversible encryption component, and a parameter recognition component.
[0034] The statistical analysis component extracts data to be analyzed from the raw data based on the parsed business application requests. It extracts preset statistical features from the data to be analyzed, resulting in extracted statistical features. These preset statistical features include numerical features, categorical features, and correlation indicators between features. Specifically, numerical features include mean, median, standard deviation, quartiles, and distribution type; categorical features include frequency distribution, mode, and cardinality; and correlation indicators include correlation coefficients and mutual information between features. It should be noted that all extracted statistical features are organized in a structured format to fully describe the statistical characteristics of the data to be analyzed. The statistical analysis component provides a foundation for the subsequent encryption process by performing comprehensive multi-dimensional statistical analysis on the data. Furthermore, the component supports incremental updates and distributed computing during the statistical analysis process, enabling efficient processing of large-scale datasets.
[0035] A reversible encryption component is used to reversibly encrypt data based on extracted statistical features to obtain encrypted data. Specifically, the process of reversibly encrypting data based on extracted statistical features to obtain encrypted data is as follows: For numerical data, a linear transformation is performed on the data based on the extracted statistical features, and differential privacy noise is added to ensure that the privacy of individual data records in the data is not leaked; the differential privacy noise is calculated based on the extracted statistical features. For categorical data, k-anonymization is performed on the data based on the extracted statistical features to map the specific values of the data to equivalence classes, thereby hiding individual differences. For textual data, calibration noise is added to the word embedding space of the data to protect the privacy of the text content; the calibration noise is calculated based on the extracted statistical features. In particular, during the reversible encryption process using the reversible encryption component, because the extracted statistical features are strongly correlated with the data, each encryption process dynamically changes, ensuring that the original data cannot be deduced from the encrypted data, but effectively maintaining the consistency of statistical characteristics.
[0036] Optionally, this embodiment 1 also provides another implementation method, specifically: replacing the reversible encryption component with a homomorphic encryption component; the homomorphic encryption component is used to encrypt the data to be analyzed based on the extracted statistical features using homomorphic encryption technology to obtain encrypted data; homomorphic encryption technology allows direct calculation on the ciphertext, but the computational complexity is high, and it is suitable for batch processing scenarios with low real-time requirements.
[0037] The parameter recognition component is used to identify parameters in the parsed processing instructions and obtain parameter recognition results; analyze the semantic context of the parameters in the parsed processing instructions to obtain the parameter data type; it uses predefined parameter data types, such as Real(value) or Sandbox(value), to determine the parameter data type; and generates instruction parameter recognition results based on the parameter data type and parameter recognition results. The process of generating instruction parameter recognition results based on parameter data type and parameter recognition results is as follows: when the parameter data type is a real business value, the parameter recognition result is directly output as the instruction parameter recognition result; when the parameter data type is a statistical inference value, a pre-calculated statistical mapping table is used to map the parameter recognition result to the real value range, and the instruction parameter recognition result is output. It is worth noting that the parameter recognition component is crucial for achieving accurate local data manipulation. When the large language model returns processing instructions after analyzing encrypted data, the parameter recognition component analyzes the semantic context of the parameters in the instructions to determine whether the parameters are real business values or statistical inferences based on statistical features. For real business values, they are used directly. For statistical inferences based on statistical features, a pre-computed statistical mapping table is queried to map them back to the range of real values, ensuring the accuracy of local data manipulation. Optionally, parameter recognition can enhance contextual semantic analysis capabilities, using natural language processing technology to more accurately determine parameter types and conversion requirements.
[0038] The instruction parsing module works closely with the MCP privacy protection tool, supporting multiple instruction formats, such as parsing natural language instructions and structured query language, and accurately extracting operation type, target data, conditional expressions, and parameter information. Optionally, the instruction parsing module also includes an instruction verification function to check the syntactic correctness and semantic rationality of the processing instructions, ensuring that the processing instructions can be executed correctly. The parsed processing instructions are transmitted to the MCP privacy protection tool in a predefined standardized format to ensure the smooth progress of subsequent processing.
[0039] The local data processing engine receives processing instructions from the MCP privacy protection tool and performs data operations such as data filtering, aggregation calculation, correlation analysis, or feature engineering on the data to be analyzed in a local environment. Optionally, the local data processing engine supports multiple data formats and data sources, such as processing structured data, semi-structured data, and unstructured data. To improve processing efficiency, the local data processing engine includes an operation optimization component, which analyzes the distribution characteristics, storage format, and indexing of the data, automatically selects the optimal execution strategy, and caches frequently used operation results to reduce redundant calculations.
[0040] Example: As attached Figure 3As shown, taking a bank anti-fraud system scenario as an example, the privacy-preserving data processing system for large language models described in Embodiment 1 will be illustrated as follows: The bank's risk control system initiates a transaction analysis request, triggering the MCP privacy protection tool in the local MCP Server layer to extract the raw transaction data from the local transaction database for the past thirty days. It then calculates statistical characteristics of the transaction amount, including mean, standard deviation, and P90 quantile, as well as the distribution characteristics of transaction time and merchant risk distribution. A reversible encryption component performs reversible encryption on the raw transaction data based on the extracted statistical characteristics. Specifically, linear transformation and noise injection are applied to sensitive numerical data such as transaction amounts, while k-anonymization is used for categorical data such as merchant information. The encrypted data is then sent to the large model service layer. At this point, the large language model in the large model service layer cannot access the raw transaction records, but can only access the statistical data. The system analyzes encrypted data and identifies fraud patterns involving high transaction amounts, high-risk merchants, and unusual times. Based on the statistical distribution characteristics of transaction amounts, the large language model infers the high-risk threshold as the P90 quantile and returns a processing instruction. After receiving the processing instruction, the parameter recognition component analyzes the semantic context of the threshold parameter in the instruction, identifies the value inferred based on statistical features, and queries the statistical mapping table to convert the P90 quantile into a true threshold of 1,750 yuan. The local data processing engine uses the converted threshold to perform filtering operations on the original transaction data. All operations are completed locally, and the original data never leaves the bank's local environment.
[0041] The privacy-preserving data processing system for large language models described in Embodiment 1 achieves a technological breakthrough that protects privacy while maintaining data availability, significantly improving key indicators such as privacy protection strength, data availability, and system performance. Specifically, the reversible encryption technology of the MCP privacy protection tool ensures that the large model can only access the raw data encrypted based on statistical features. The raw data is always processed locally, completely eliminating the risk of data leakage and meeting the stringent data security requirements of highly sensitive industries such as finance and healthcare. Specifically, the parameter recognition component identifies parameters in the parsed processing instructions, ensuring that business decisions made by the large model based on encrypted data are accurately executed on the locally stored raw data, maintaining the accuracy and reliability of business decisions. Furthermore, the localized processing architecture design ensures that all data operations are completed locally, reducing the amount of data transmitted over the network, improving processing efficiency, and supporting the needs of real-time business scenarios.
[0042] In this embodiment 1, the user environment layer is used to store the original data locally and access business application requests, ensuring the physical isolation of the original data. The local MCP Server layer is used for statistical feature extraction, reversible encryption, and parameter identification in the processing instructions issued by the large model, achieving a balance between privacy protection and data availability. Specifically, by extracting preset statistical features from the data to be analyzed, the extracted statistical features describe the overall characteristics of the data rather than individual information, eliminating the need for direct processing of the original data and fundamentally avoiding the risk of privacy leakage. Encryption is performed based on the extracted statistical features, ensuring that the encrypted data retains its statistical characteristics, allowing the large model to still perform effective analysis. By introducing a mechanism of real business values and statistical inference values, when the large model returns processing instructions, it accurately identifies whether the parameters are real business values or statistical inference values. For statistical inference values, a statistical mapping table is used to convert them into the range of real values, ensuring accurate execution on the original data. The encrypted data is analyzed in the large model service layer, and processing instructions are returned, which are then processed by the local MCP. The server layer performs data manipulation operations on the raw data, while the large model service layer only receives privacy-protected data and cannot access the raw data, thus eliminating the risk of raw data leakage and ensuring an effective balance between data privacy protection and large model analysis capabilities.
[0043] Example 2 This embodiment 2 provides a privacy-preserving data processing method for large language models, utilizing the privacy-preserving data processing system for large language models provided in embodiment 1 above. (See attached...) Figure 4 As shown, the privacy-preserving data processing method for large language models includes the following steps: Step 1: Store the raw data locally and receive requests from business applications.
[0044] Step 2: Extract the data to be analyzed from the raw data according to the business application request; encrypt the data to be analyzed according to the pre-extracted statistical features to obtain encrypted data; and use it to perform preset data operation processing on the data to be analyzed based on the processing instructions issued by the large language model.
[0045] Step 3: Analyze the encrypted data using a large language model and issue processing instructions.
[0046] It should be noted that the specific process of the privacy-preserving data processing method for large language models described in Embodiment 2 is described in the corresponding content of Embodiment 1 above, and will not be repeated here.
[0047] The privacy-preserving data processing system for large language models described in this invention is based on a statistical feature-driven privacy protection concept. It fundamentally changes the traditional privacy protection approach by first extracting statistical features from the data and then performing privacy protection based on these features. This effectively resolves the fundamental contradiction between large model analysis capabilities and data privacy protection. Specifically, the statistical feature-based privacy protection mechanism, through the MCP protocol tool, achieves collaborative work of data statistical feature extraction, reversible encryption processing, and parameter identification processing, resolving the fundamental contradiction between large models and data privacy. It achieves a balance between privacy protection and usability, maintaining data statistical characteristics while protecting data privacy, enabling effective analysis of large models. Simultaneously, it enables real-time data processing capabilities, meeting the technical requirements of actual business scenarios.
[0048] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.
Claims
1. A privacy-preserving data processing system for large language models, characterized in that, This includes the user environment layer, the local MCP Server layer, and the large model service layer; The user environment layer is used to store raw data locally and receive requests from business applications. The local MCP Server layer is used to extract the data to be analyzed from the raw data according to the business application request; encrypt the data to be analyzed according to the pre-extracted statistical features to obtain encrypted data; and perform preset data operation processing on the data to be analyzed based on the processing instructions issued by the large language model. The large model service layer is used to analyze encrypted data using a large language model and issue processing instructions to the local MCPServer layer.
2. The privacy-preserving data processing system for large language models according to claim 1, characterized in that, The local MCP Server layer includes an instruction parsing module, MCP privacy protection tools, and a local data processing engine; The instruction parsing module is used to receive and parse business application requests to obtain the parsed business application requests. It is also used to receive and parse the processing instructions issued by the large language model, and to pass the parsed processing instructions to the MCP privacy protection tool; The MCP privacy protection tool is used to extract data to be analyzed from the raw data based on the parsed business application requests, and to extract statistical features from the data to be analyzed; based on the extracted statistical features, the data to be analyzed is encrypted to obtain encrypted data; it is also used to identify parameters of the parsed processing instructions to obtain instruction parameter identification results. The local data processing engine is used to identify the results based on the command parameters and perform preset data operations on the data to be analyzed.
3. The privacy-preserving data processing system for large language models according to claim 2, characterized in that, MCP privacy protection tools include statistical analysis components, reversible encryption components, and parameter recognition components; The statistical analysis component is used to extract the data to be analyzed from the raw data based on the parsed business application requests, extract the preset statistical features of the data to be analyzed, and obtain the extracted statistical features. A reversible encryption component is used to reversibly encrypt the data to be analyzed based on extracted statistical features to obtain encrypted data. The parameter recognition component is used to identify the parameters in the parsed processing instructions and obtain the parameter recognition results; The semantic context of the parameters in the parsed processing instructions is analyzed to obtain the parameter data type; Based on the parameter data type and parameter recognition results, generate instruction parameter recognition results.
4. A privacy-preserving data processing system for large language models according to claim 3, characterized in that, The preset statistical characteristics of the data to be analyzed include numerical characteristics, categorical characteristics, and correlation indicators between characteristics; Numerical characteristics include mean, median, standard deviation, quartiles, and distribution type; Categorical features include frequency distribution, mode, and cardinality; The correlation indicators of features include the correlation coefficient and mutual information between features.
5. A privacy-preserving data processing system for large language models according to claim 3, characterized in that, Based on the extracted statistical features, the data to be analyzed is reversibly encrypted to obtain encrypted data, as follows: For numerical data to be analyzed, a linear transformation is performed on the data based on the extracted statistical features, and differential privacy noise is added; the differential privacy noise is calculated based on the extracted statistical features. For categorical data to be analyzed, k-anonymization is performed on the data based on the extracted statistical features. For text-based data to be analyzed, calibration noise is added to the word embedding space of the data; the calibration noise is calculated based on the extracted statistical features.
6. A privacy-preserving data processing system for large language models according to claim 3, characterized in that, The process of generating instruction parameter recognition results based on parameter data type and parameter recognition results is as follows: When the parameter data type is a real business value, the parameter recognition result will be directly output as the instruction parameter recognition result; When the parameter data type is a statistical inference value, the parameter identification result is mapped to the true value range using a pre-calculated statistical mapping table, and the instruction parameter identification result is output.
7. A privacy-preserving data processing system for large language models according to claim 2, characterized in that, The preset data processing operations include data filtering, aggregation calculation, correlation analysis, and feature engineering.
8. A privacy-preserving data processing system for large language models according to claim 2, characterized in that, The instruction parsing module supports parsing natural language instructions and structured query languages; the local data processing engine supports processing structured data, semi-structured data, and unstructured data.
9. A privacy-preserving data processing system for large language models according to claim 1, characterized in that, The raw data is financial transaction data, government data, or medical data.
10. A privacy-preserving data processing method for large language models, characterized in that, Using the privacy-preserving data processing system for large language models as described in any one of claims 1-9; Store the raw data locally and receive requests from business applications; Based on business application requests, extract the data to be analyzed from the raw data; encrypt the data to be analyzed based on the pre-extracted statistical features to obtain encrypted data; and use it to perform preset data operation processing on the data to be analyzed based on the processing instructions issued by the large language model. The encrypted data is analyzed using a large language model, and processing instructions are issued.