Reexamining anonymity claims of synthetic data generation: a model-centric privacy attack perspective

A study argues that the anonymity of synthetic data generation should be reevaluated from a model-centric privacy attack perspective and cannot be judged based on a single dataset alone. Based on the GDPR definition, the study maps identifiability risk to privacy attacks under different threat settings, stating that synthetic data alone is insufficient to guarantee anonymization; after comparing DP and SBPMs, the authors argue that DP can provide robust protection against identifiability risk, while SBPMs lack sufficient safeguards.