CN-122020693-A - Agent-based large model dynamic desensitization system and method

CN122020693ACN 122020693 ACN122020693 ACN 122020693ACN-122020693-A

Abstract

The invention discloses an Agent-based large model dynamic desensitization system and method, wherein the system comprises a multi-source perception module, an intelligent decision module, a large model semantic enhancement module and a dynamic desensitization execution module, wherein the multi-source perception module is used for acquiring and structuring multi-dimensional information in real time, the intelligent decision module is used for generating a self-adaptive desensitization strategy based on the multi-dimensional information through a decision mechanism fused by a rule engine and a reinforcement learning model, the large model semantic enhancement module is used for carrying out semantic understanding on input data by using a large model which is accessed in advance, identifying sensitive contents including dominant sensitive information and contextual implicit sensitive information, and the dynamic desensitization execution module is used for executing desensitization operation on the identified sensitive contents through a protocol self-adaption and data length compensation mechanism according to the desensitization strategy, so that the problems of poor compatibility, insufficient suitability, lack of self-adaption decision capability and passive rule execution in the existing dynamic desensitization technology are solved.

Inventors

CHEN KAIPING
Hui Lulu
Zhong Zixuan
XIE SHUHANG

Assignees

杭州安泉数智科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. An Agent-based large model dynamic desensitization system, comprising: the multi-source sensing module is used for collecting and structuring multi-dimensional information in real time; The intelligent decision module is connected with the multi-source perception module and is used for generating a self-adaptive desensitization strategy based on the multi-dimensional information through a decision mechanism fused by a rule engine and a reinforcement learning model; the large model semantic enhancement module is connected with the intelligent decision module and is used for carrying out semantic understanding on input data by using a large model which is accessed in advance and identifying sensitive contents including explicit sensitive information and contextual implicit sensitive information; The dynamic desensitization execution module is respectively connected with the intelligent decision module and the large model semantic enhancement module and is used for executing desensitization operation on the identified sensitive content through a protocol self-adaption and data length compensation mechanism according to the desensitization strategy.
2. The Agent-based large model dynamic desensitization system according to claim 1, wherein the multidimensional information comprises information of user identity, application scene, data type and environment state dimension, and the user identity dimension information comprises emotion affinity tags generated based on user historical interaction behavior analysis.
3. The Agent-based large model dynamic desensitization system according to claim 1, wherein the intelligent decision module comprises a rule engine module, a reinforcement learning module and a conflict mediation module, wherein the rule engine module is internally provided with industry general rules and user-defined rules, the reinforcement learning module carries out policy optimization on three-dimensional targets formed by desensitization safety, data availability and system performance loss, and the conflict mediation module is used for outputting a final policy according to preset priority rules and real-time risk quantification results when policies generated based on different dimension information conflict.
4. The Agent-based large model dynamic desensitization system according to claim 1, wherein the large model pre-accessed in the large model semantic enhancement module is a model tuned by a low-rank adaptation technique, wherein the tuning comprises: Freezing the original parameters of the large model; Injecting a trainable low-rank decomposition matrix pair beside the self-attention module linear projection weight of the transducer layer of the large model; The low rank decomposition matrix is trained using the sensitive information identification annotation dataset.
5. The Agent-based large model dynamic desensitization system according to claim 1, wherein said large model semantic enhancement module is identified by named entity identification and sensitivity scoring algorithm, specifically comprising: obtaining a candidate entity set through named entity identification; assigning basic sensitivity scores for each entity according to the domain knowledge base; analyzing the context co-occurrence relationship among the entities to calculate the association weight; and determining the context implicit sensitive information according to the final sensitivity score and a preset threshold value.
6. An Agent-based large model dynamic desensitization system according to claim 1, wherein said data length compensation mechanism is implemented by creating desensitized views that are exactly identical to the original data field lengths.
7. The Agent-based large model dynamic desensitization system according to claim 1, wherein said protocol adaptation specifically comprises: carrying out protocol analysis on an original data stream from a client at an application gateway layer to extract a data load; The data load is sent to a dynamic desensitization execution module for processing; and reorganizing the processed data into a new data stream according to the original protocol specification and returning the new data stream.
8. The Agent-based large model dynamic desensitization system according to claim 1, further comprising a strategy optimization module, wherein the strategy optimization module is connected with the dynamic desensitization execution module and the intelligent decision module and is used for collecting desensitization effect feedback, and performing iterative optimization on the reinforcement learning model and the large model semantic enhancement module based on the feedback to form a closed-loop desensitization strategy.
9. The Agent-based large model dynamic desensitization system according to claim 1, wherein the dynamic desensitization execution module supports a reversible desensitization mode, the intelligent decision module is further configured to evaluate risk levels of application scene dimensions according to trust in user identity dimensions, dynamically decide whether to enable the reversible desensitization mode for specific sensitive content, and when enabled, the reversible desensitization mode performs desensitization by using encryption parameters bound with current users and scenes, and generates corresponding traceable authorization metadata.
10. An Agent-based large model dynamic desensitization method is characterized by comprising the following steps: collecting and structuring multidimensional information in real time; based on the multidimensional information, generating a self-adaptive desensitization strategy according to a decision mechanism fused by a rule engine and a reinforcement learning model; carrying out semantic understanding on input data by using a pre-accessed large model, and identifying sensitive contents including explicit sensitive information and context implicit sensitive information; Based on the desensitization strategy, the desensitization operation is carried out on the identified sensitive content through a protocol self-adaption and data length compensation mechanism.

Description

Agent-based large model dynamic desensitization system and method Technical Field The invention relates to the technical field of data security processing, in particular to an Agent-based large model dynamic desensitization system and method. Background In recent years, large models are widely used in various industries. During large model interactions, the data entered by users often contains large amounts of sensitive information, such as personal identity information, business secrets, medical records, etc., which, once compromised, may pose privacy risks and economic losses. Traditional data desensitization technology is mainly based on static rules or simple pattern matching, and is difficult to adapt to complexity and dynamics of large-model interaction scenes. In recent years, partial researches try to introduce a machine learning mechanism to construct a dynamic desensitization scheme, but in practice, a plurality of challenges still exist, namely, some schemes realize desensitization by rewriting SQL sentences or modifying a database return result, are required to be adapted to a specific database protocol, are limited by the sealing performance of a commercial database protocol, the universality is restricted, the data length change after the desensitization can influence the transmission stability, additional performance expenditure is brought, other middleware schemes have stronger dependence on a client connection mode, the deployment difficulty is increased when a third party application or a stock system is docked, the application range is limited, in addition, most of the existing methods do not fully fuse multidimensional information such as user identities, interaction contexts, scene risks and the like to carry out strategy generation, mismatching of desensitization granularity and actual requirements is easy to cause the influence on the usability or the safety protection effect of data, meanwhile, the current desensitization mechanism mostly adopts a fixed rule driving mode, lacks active perception and intelligent decision of an interaction process and continuous optimization capability based on feedback, and has insufficient support for diversified sensitive data types and complex interaction scenes. Disclosure of Invention In view of the above, the present invention aims to provide a large model dynamic desensitization system and method based on Agent, which are used for solving the technical problems of poor compatibility, insufficient adaptability, lack of adaptive decision capability, passive rule execution and the like existing in the existing dynamic desensitization technology, and realizing the effects of wide compatibility, balanced desensitization precision and availability, and strong scene adaptability, and the specific scheme is as follows: in a first aspect, the present application provides an Agent-based large model dynamic desensitization system, comprising: the multi-source sensing module is used for collecting and structuring multi-dimensional information in real time; The intelligent decision module is connected with the multi-source perception module and is used for generating a self-adaptive desensitization strategy based on the multi-dimensional information through a decision mechanism fused by a rule engine and a reinforcement learning model; the large model semantic enhancement module is connected with the intelligent decision module and is used for carrying out semantic understanding on input data by using a large model which is accessed in advance and identifying sensitive contents including explicit sensitive information and contextual implicit sensitive information; The dynamic desensitization execution module is respectively connected with the intelligent decision module and the large model semantic enhancement module and is used for executing desensitization operation on the identified sensitive content through a protocol self-adaption and data length compensation mechanism according to the desensitization strategy. As a preferable technical scheme of the invention, the multidimensional information comprises information of user identity, application scene, data type and environment state dimension, and the user identity dimension information comprises emotion affinity tags generated based on user historical interaction behavior analysis. The intelligent decision module comprises a rule engine module, a reinforcement learning module and a conflict adjustment module, wherein the rule engine module is internally provided with industry general rules and user-defined rules, the reinforcement learning module carries out policy optimization on a three-dimensional target formed by desensitization safety, data availability and system performance loss, and the conflict adjustment module is used for outputting a final policy according to a preset priority rule and a real-time risk quantification result when policies generated based on different dimensional information