A generative artificial intelligence user data management and control method

Abstract

The application discloses a generative artificial intelligence user data management and control method, adopts input management and control software, output management and control software and silent management and control software to form a software management and control architecture, realizes pre-safety detection, output safety detection and response violation detection through text feature matching, realizes interface fine silent disablement in combination with a storage permission index table, does not generate an execution log in a management and control process, can be matched with a setting isolated file box, a user notification link and a safety management and control report generation function, effectively avoids secondary diffusion of sensitive information, and is suitable for generative artificial intelligence safety management and control scenes of various organizations and institutions.

Description

1. Technical Field This invention belongs to the field of artificial intelligence information security technology, specifically relating to a generative artificial intelligence user data security management method based on the collaboration of multiple software modules. 2. Background Technology Existing generative AI management and control systems mostly use single management and control modules, with security control for command input, data output, and model feedback being fragmented and failing to form a collaborative management and control system. Violation data lacks a unified handling mechanism, which can easily lead to the secondary spread of sensitive information. Furthermore, they do not combine user identity identification to achieve refined access control, resulting in insufficient security effectiveness. At the same time, existing management and control solutions often retain violation management and control execution logs, posing secondary security risks caused by log leakage. 3. Summary of the Invention 3.1 Terminology Definition Storage permission index table: A structured data table built during system initialization, with the user identity identifier as the primary key, used to map the association between the user identity identifier and the permissions of the large model software logic output interface; Input control software: Deployed at the system entry layer, this software module is used to perform pre-security checks on user business commands and intercept illegal commands. Output control software: Deployed in the middle layer of the system, it is a software module used to perform output security checks on command data and block illegal data; Silent control software: Deployed at the system egress layer, this software module is used to detect violations in the response results returned by large models, disable interface permissions, and not generate execution logs. Software management architecture: A full-chain software security management system composed of input management software, output management software, and silent management software working together in sequence; Isolated File Box: A dedicated storage space within the system that features independent storage and prohibits external access, used to store illegal commands, illegal data, and illegal response results; Large Model Software Logic Output Interface: The software logic interface used by the generative artificial intelligence large model to transmit response results to the outside world; User Identifier: Character information used to uniquely identify the user performing the operation; Security control report: An audit report that records control operations such as system violation detection, interface disabling, and log generation.

[0001] 3.2 Technical Solution This invention employs a software management architecture comprised of input management software, output management software, and silent management software. Combined with text feature matching and detection technology, it achieves end-to-end security management of generative artificial intelligence user data. The specific steps are as follows: Upon system power-on initialization, a storage permission index table with the user's identity as the primary key is built in the system database; The input control software receives user business instructions and performs pre-security checks using a text feature matching algorithm commonly used in the field. Specifically, the pre-security checks detect whether the business instructions carry illegal field tampering identifiers, identify and intercept illegal instructions carrying illegal field tampering identifiers, generate system operation logs and archive them according to system preset rules, move illegal instructions associated with user identity identifiers into an isolated file box, and transfer compliant instructions to the output control software. The output control software uses a text feature matching algorithm commonly used in the field to perform output security detection on compliant instructions. Specifically, the output security detection detects whether the instruction contains illegal privacy data call fields, identifies and blocks illegal data containing illegal privacy data call fields, generates system operation logs and archives them according to the system's preset rules, moves illegal data associated with user identity identifiers into an isolated file box, and transmits compliant instructions to a generative artificial intelligence big model. Generative AI large-scale models generate response results based on compliant instructions and send them back to the silent management software; The silent management software uses a text violation recognition model commonly used in the field to perform violation detection of response results. Specifically, the violation detection is to detect whether the response result is inducing the user to disclose sensitive privacy information. After identifying the violation response result that induces the user to disclose sensitive privacy information, the software locates the corresponding user's large model software logic output interface based on the storage permission index table, and marks the interface permission status as 0 to disable it. This management operation does not generate any execution logs, and moves the violation response result associated with the user's identity to an isolated file box. The system can inform users of the results of this violation control through the user terminal interface, and can also generate a security control report based on the system operation log to complete the full-link security control.

[0002] 3.3 Beneficial Effects 1. A multi-software module collaborative system is adopted to form a full-link control system. Through text feature matching detection rules, precise and secure control is achieved in all aspects of instruction input, data output, and model feedback. 2. Implement fine-grained user-level permission control based on the storage permission index table, significantly improving the accuracy of control; 3. By isolating files and collecting prohibited content, we can prevent the secondary spread of sensitive information from the source. 4. The silent control process controls the output information of the generative artificial intelligence large model, does not generate execution logs for the control operation, and does not send control execution records to users, thus completely eliminating the secondary security risks caused by the leakage of logs and execution records; 5. Add a user terminal notification step to protect users' right to know and improve the rationality of system interaction; 6. Security control reports can be generated based on system operation logs, facilitating system operation and maintenance and security auditing; 7. The core detection logic is implemented using a common detection algorithm in this field. The technical solution is highly versatile and can be adapted to generative artificial intelligence security management scenarios of various organizations. 4. Detailed Implementation Example

[0003] A senior user of a commercial bank initiated a personal business inquiry command through the bank's intelligent AI terminal. The command entered the bank's AI software control architecture via an encrypted communication link. The input control software uses a text feature matching algorithm to perform pre-security checks on business instructions, identify and intercept illegal instructions that carry illegal field tampering identifiers, generate system operation logs and archive them according to the bank's system rules, move the illegal instructions associated with user identity identifiers into the system isolation file box, and transfer the remaining compliant instructions to the output control software; The output control software uses a text feature matching algorithm to perform security checks on the execution of compliant instructions, identifies and blocks illegal data containing fields that violate privacy data access rules, generates system operation logs and archives them according to the bank's system rules, moves the illegal data to the system's isolated file box along with the user's identity identifier, and transmits the remaining compliant instructions to the bank's generative artificial intelligence big data model. Generative AI large-scale models generate response results based on compliant instructions and send them back to the silent management software; The silent management software uses a general text violation recognition model to detect violations that lead users to disclose sensitive privacy information. Based on the storage permission index table, it locates the large model software logic output interface corresponding to the user and marks the interface permission status as 0 to disable it. This management operation does not generate an execution log and moves the violation response result, associated with the user's identity, into the system's isolated file box. The system informs the user of the violation control result through the bank's intelligent AI terminal interface, and generates a security control report based on the system operation log, thus completing the full-link security control.

Claims

1. A generative artificial intelligence user data management and control method, characterized in that, A software management architecture consisting of input management software, output management software, and silent management software is adopted, including: (1) The input control software performs pre-security checks on user business instructions. Specifically, the pre-security checks detect whether the business instructions carry illegal field tampering identifiers, intercept the non-compliant instructions carrying illegal field tampering identifiers, generate system operation logs, and transfer compliant instructions to the output control software. (2) The output control software performs output security detection on the compliant instructions. Specifically, the output security detection detects whether the instructions contain illegal privacy data call fields, blocks illegal data containing illegal privacy data call fields, generates system operation logs, and transmits the compliant instructions to the generative artificial intelligence big model. (3) The silent control software performs violation detection on the response results returned by the generative artificial intelligence large model. Specifically, the violation detection is to detect whether the response results lead the user to disclose sensitive privacy information, disable the large model software logic output interface corresponding to the situation, and not generate the execution log of the control operation.

2. The method according to claim 1, characterized in that, It also includes system initialization steps, constructing a storage permission index table with the user's identity identifier as the primary key, and the silent management software locating and executing interface permission disabling operations based on the storage permission index table.

3. The method according to claim 1, characterized in that, The system operation logs are archived and retained according to the system's preset rules.

4. The method according to claim 1, characterized in that, The input control software will associate the violation commands with the user's identity and move them to an isolated file box.

5. The method according to claim 1, characterized in that, The output control software will associate the violation data with the user's identity and move it into an isolated file box.

6. The method according to claim 1, characterized in that, The silent management software moves the user identity identifier associated with the violation response results into an isolated file box.

7. The method according to claim 1, characterized in that, The large model software logic output interface is disabled by setting the permission status to 0.

8. The method according to claim 1, characterized in that, It also includes the step of informing users of the results of violation control through the user terminal interface.

9. The method according to claim 1, characterized in that, It also includes the step of generating a security control report based on the system operation log.