Model Inversion Reconstructing Training Data
模型逆向重建訓練資料
Design & Development
Risk Description
When an attacker conducts extensive queries against a model trained on sensitive imagery and reconstructs recognizable approximations of the originals; due to unmitigated control gaps, the organization assumed the model itself contained no source data and had no defenses against this risk, triggering external stakeholder impacts and causing the incident constitutes a substantive breach.
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
Train models on sensitive imagery with differential privacy to reduce single-sample reconstructability; Apply output perturbation, query rate limits and anomalous-query detection at the inference interface to raise reconstruction cost; Include model-inversion attacks in pre-launch red-team testing