S0036RS3-T08-S0036-Z · Full risk code

Label Tampering Attack

標籤竄改攻擊

Design & Development
Risk Description

When an outsourced annotated dataset is maliciously altered in a data poisoning attack; because pre-deployment validation checks only aggregate accuracy without label integrity or gold-sample audits, the model consistently misclassifies boundary categories, breaching safety defenses and incurring retraining costs.

Framework Mappings

OWASP Top 10 for LLMLLM04
MITRE ATLASAML.T0020
NIST AI 100-2Poisoning
ISO/IEC 5338設計與開發
MIT AI Risk RepositoryDomain 2

Risk Treatment & Implementation Guidance

Verify outsourced labeling integrity with gold-sample injection and cross-annotation audits; Maintain provenance tagging locating each label batch's provider and version; Include per-class and boundary-case review in validation rather than overall accuracy alone