S0106RS3-T06-S0106-Z · Full risk code

Data Leakage Through Model Memorization

模型記憶化導致資料外洩

Design & Development
Risk Description

When an attacker repeatedly tests specially formatted prompts and succeeds in extracting real personal information from the training data through model output; due to unmitigated control gaps, the organization believed the data had been adequately processed but never assessed memorization risk, triggering external stakeholder impacts and causing the incident constitutes a personal data breach requiring notification.

Framework Mappings

OWASP Top 10 for LLMLLM02
NIST AI 600-1Data Privacy
MITRE ATLASAML.T0024
ISO/IEC 5338運作與監控
MIT AI Risk RepositoryDomain 2

Risk Treatment & Implementation Guidance

Assess and mitigate memorization risk in training, applying differential privacy or equivalent techniques when personal data is involved; Run memorization-detection tests before launch simulating prompts that elicit training data; Filter suspected verbatim personal data at the output layer to block leakage