An AI training data obligation reasoning method and system based on a training obligation axiom library extension and a storage medium
Patent Information
- Application Number
- CN202610941267.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-28
- Publication Date
- 2026-09-25
AI Technical Summary
针对现有训练数据义务不被推理、无整体义务剖面的问题，本发明提供一种基于训练义务公理库扩展的AI训练数据义务推理方法、系统及存储介质
其一，以TBox-AI推理批量训练数据义务、形成整体义务剖面；其二，聚合取最严、就严不就宽；其三，是安全义务内生方法在AI训练场景的扩展，为训练数据义务凭证签发与跨模型反向追溯方法 凭证签发供依据（已完整工程设计）。
Smart Images
Figure CN122819459A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of data security and artificial intelligence, and in particular to a method, system, and storage medium for training data security obligations and calculating the most stringent obligation profile using inference AI. Background Technology
[0002] AI model training uses a large amount of training data, each of which carries a safety obligation. The trained model should be bound by these obligations, therefore it is necessary to infer the obligations of the batch training data and form an overall obligation profile.
[0003] Current practices often fail to infer the obligations of training data or fail to form the strictest obligation profile of the entire training set, causing the model to deviate from the obligation constraints of the training data. Therefore, how to infer the obligations of batch training data using an expanded training obligation axiom library and aggregate the strictest obligations to form an obligation profile is a key issue in AI training data obligation governance. Summary of the Invention
[0004] (a) Technical problems to be solved To address the problems of existing training data obligations not being inferred and the lack of an overall obligation profile, this invention provides an AI training data obligation inference method, system, and storage medium based on an extended training obligation axiom library.
[0005] (II) Technical Solution This invention extends the training obligation axiom library (TBox-AI) on the general obligation axiom library; it infers obligations one by one from batch training data; and it aggregates and takes the most stringent one to form the most stringent obligation profile. The training data obligations are derived from the data ontology (an AI extension of the security obligation endogenous method), and the most stringent obligation profile is used for issuing training data obligation certificates and cross-model backtracking methods to issue training data obligation certificates.
[0006] (III) Beneficial Effects First, TBox-AI inference batch training data obligations are used to form an overall obligation profile; second, the strictest requirement is aggregated and the strictest requirement is adopted; third, it is an extension of the security obligation endogenous method in AI training scenarios, providing a basis for the issuance of training data obligation certificates and cross-model reverse traceability method certificates (complete engineering design has been completed). Attached Figure Description
[0007] Figure 1 This is a schematic diagram of AI training data obligation inference based on the training obligation axiom library extension.
[0008] Figure 2 A schematic diagram for training the Obligation Axiom Library (TBox-AI).
[0009] Figure 3 This is a schematic diagram of batch inference and the strictest obligation profile.
[0010] Figure 4 Flowchart of inference task for training data.
[0011] Figure 5 This diagram illustrates the connection between the intrinsic method of security obligations, the issuance of training data obligation certificates, and the cross-model reverse tracing method. Detailed Implementation
[0012] The present invention will be further described in detail below with reference to the accompanying drawings, using a profile of an inference task on a batch of training data as an example.
[0013] 1. Overall Mechanism See Figure 1 The extended training obligation axiom library is used to infer the obligation of batch training data and aggregate the most stringent obligation profile.
[0014] 2. Training Obligation Axiom Library See Figure 2 TBox-AI extends the obligation inference rules for AI training scenarios based on the general obligation axiom.
[0015] 3. Batch reasoning and the most stringent profile See Figure 3 The task involves reasoning about batches of data one by one, aggregating the most stringent data to form the training set task profile, and choosing the strictest approach over the lenient one.
[0016] 4. Reasoning Process See Figure 4 It acquires batch data, performs TBox-AI inference tasks, aggregates the most stringent data, and outputs the most stringent task profile.
[0017] 5. Integration with endogenous methods for security obligations, issuance of training data obligation certificates, and cross-model reverse tracing methods. See Figure 5 This invention is an AI extension of the intrinsic security obligation method, and a method for issuing training data obligation certificates and cross-model backtracking issues training data obligation certificates based on the strictest obligation profile. The above embodiments are for illustrative purposes only and not for limitation; all equivalent substitutions and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A method for AI training data obligation inference based on an extended training obligation axiom library, characterized in that, include: A training obligation axiom library is obtained by extending the general safety obligation axiom library; the safety obligations of the batch AI training data are inferred one by one using the training obligation axiom library; the inferred data obligations are aggregated and the most stringent constraint is taken to obtain the most stringent obligation profile of the batch AI training data.
2. The method according to claim 1, characterized in that, The training obligation axiom library extends the obligation inference rules for AI training scenarios on top of the general safety obligation axiom library.
3. The method according to claim 1, characterized in that, The batch inference process performs inference tasks on batch training data one by one according to the training task axiom library.
4. The method according to claim 1, characterized in that, The aggregation of the most stringent data obligations forms the training set obligation profile by selecting the most stringent constraint among all data obligations.
5. The method according to claim 1, characterized in that, The strictest obligation profile represents the strictest security obligation that the training dataset as a whole should satisfy.
6. The method according to claim 1, characterized in that, The security obligations of the training data are derived from the data ontology, and the axiom library of training obligations is an extension of the security obligations inherent in the AI training scenario.
7. The method according to claim 1, characterized in that, The most stringent obligation profile is used for issuing training data obligation certificates.
8. The method according to claim 1, characterized in that, The batch inference supports batch processing of large-scale training data.
9. The method according to claim 1, characterized in that, The training obligation axiom library is extensible.
10. The method according to claim 1, characterized in that, The reasoning and aggregation process will be documented.
11. The method according to claim 10, characterized in that, The traces are recorded using SM3 hashing, and the signatures are recorded using SM2.
12. An AI training data obligation inference system based on an extended training obligation axiom library, characterized in that, include: The system includes an axiom library module for maintaining an extended training obligation axiom library; an inference module for inferring obligations one by one from batch training data; and an aggregation module for aggregating and taking the most stringent requirement to form an obligation profile. The system is used to execute the method of any one of claims 1 to 11.
13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 11.