|
Work item:
|
X.AA-LLM
|
|
Subject/title:
|
Guidelines for Preventing and Mitigating Adversarial Attacks on LLMs in Metaverse and Digital Twin Environments
|
|
Status:
|
Under study
|
|
Approval process:
|
AAP
|
|
Type of work item:
|
Recommendation
|
|
Version:
|
New
|
|
Equivalent number:
|
-
|
|
Timing:
|
2026-Q4 (Medium priority)
|
|
Liaison:
|
-
|
|
Supporting members:
|
-
|
|
Summary:
|
This Recommendation provides actionable guidelines to mitigate adversarial attacks on LLMs in Metaverse and Digital Twin systems, where manipulated inputs threaten system integrity, privacy, and user trust.
Key Guidelines:
Preventive Measures:
Input validation for multi-modal data (text/voice/AR).
Adversarial training with synthetic attack scenarios.
Model hardening (e.g., output constraints).
Detection & Response:
Real-time anomaly detection (transformer-based monitoring).
Automated containment of compromised instances (<50ms response).
Procedures for Mitigating Successful Attacks
Input validation
Robust training pipelines
Anamoly detection
|
|
Comment:
|
-
|
|
Reference(s):
|
|
|
Historic references:
|
|
Contact(s):
|
|
| ITU-T A.5 justification(s): |
|
|
|
|
First registration in the WP:
2025-04-17 15:15:25
|
|
Last update:
2025-04-17 15:20:15
|
