Membership Inference Revealing Participation
成員推斷攻擊揭露參與事實
Design & Development
Risk Description
When an attacker conducts extensive comparative queries against a disease prediction model and infers that a particular individual's record was included in the training set, effectively disclosing that person's health status; due to unmitigated control gaps, although the organization did not directly leak data, substantial privacy harm has occurred and is difficult to remedy, triggering compliance exposure and operational reputational costs.
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 data with differential privacy to reduce the identifiable influence of individual records; Add output perturbation or limit query granularity to raise the cost of membership inference; Monitor abnormal high-volume comparative querying and enforce rate limits