|
Work item:
|
X.rg-dis
|
|
Subject/title:
|
Requirements for guidelines for data interaction security in training and inference stages of generative Artificial Intelligence
|
|
Status:
|
Under study
|
|
Approval process:
|
TAP
|
|
Type of work item:
|
Recommendation
|
|
Version:
|
New
|
|
Equivalent number:
|
-
|
|
Timing:
|
2027-12 (Medium priority)
|
|
Liaison:
|
ITU-T SG21
|
|
Supporting members:
|
China Mobile, China Unicom
|
|
Summary:
|
This proposal addresses key challenges in GenAI, including the significant security threats and protection pressures it faces, as well as enterprises' inability to understand the internal protection logic of models when using GenAI services. Against this backdrop, it analyzes the data interaction specific to GenAI. The proposal also summarizes and examines security threats throughout GenAI’s data interaction process: those that may disrupt the model’s normal inference and operation via interactive data; those that could impair the model’s regular functionality and service delivery through abnormal interaction behaviors; and those where the model itself outputs inappropriate data, thereby impacting end-users. Drawing on data interaction methods, the proposal puts forward data interaction security requirements for guidelines. By detecting and handling interactive data, this recommendation aims to mitigate security risks posed by such data and behaviors to both models and users. Specifically, the recommendation comprises three functional modules—training data input detection and handling, inference input detection and handling, and inference output detection and handling—to manage and inspect interactive data. Additionally, a general support capability module is designed to provide foundational, cross-cutting support for the three core modules mentioned above.
|
|
Comment:
|
-
|
|
Reference(s):
|
|
|
Historic references:
|
|
Contact(s):
|
|
| ITU-T A.5 justification(s): |
|
|
|
|
First registration in the WP:
2025-12-11 13:18:58
|
|
Last update:
2025-12-11 13:21:02
|
