|
Work item:
|
X.scm-llm
|
|
Subject/title:
|
Security control methods for data of large language models
|
|
Status:
|
Under study
|
|
Approval process:
|
TAP
|
|
Type of work item:
|
Recommendation
|
|
Version:
|
New
|
|
Equivalent number:
|
-
|
|
Timing:
|
2028-Q1 (Medium priority)
|
|
Liaison:
|
ITU-T SG21, ISO/IEC JTC 1/SC27/WG4 and WG5, ISO/IEC JTC 1/SC42, ETSI TC SIA
|
|
Supporting members:
|
China Mobile, China Academy of Information and Communications Technology
|
|
Summary:
|
Large language models are characterized by a huge parameter scale, complex internal logic, and high dependence on massive high-quality data. They may have internal defects, including memorization effects, overfitting, and insufficient adversarial training. LLMs have been widely used in various fields to receive and process massive amounts of information. Affected by the characteristics and internal defects of LLMs, security threats such as data leakage, data theft and data integrity issues may arise during model development and deployment. This draft Recommendation provides the data security threats of LLMs, and puts forward corresponding guidelines for risk identification and security control methods. It can guide model developers and users to identify data security threats of LLMs and implement corresponding security controls, so as to reduce the occurrence probability of data security incidents during the application of large language models.
|
|
Comment:
|
-
|
|
Reference(s):
|
|
|
Historic references:
|
|
Contact(s):
|
|
| ITU-T A.5 justification(s): |
|
|
|
|
First registration in the WP:
2026-04-02 14:27:47
|
|
Last update:
2026-06-10 16:08:57
|
