Exploring The Empirical Privacy Landscape of Synthetic Data
Rewriting records to replace unique private details improved the balance between model capability and measured privacy leakage.
The researchers compared methods using synthetic medical records seeded with canaries to detect leakage. Direct fine-tuning delivered the strongest capabilities but leaked the most; DP-SGD reduced leakage at a substantial capability cost. Privacy-aware rewriting protected information better than naive paraphrasing across multiple rewriting models. More capable rewriters generally reduced downstream leakage, but extra test-time compute produced inconsistent gains.
HF Daily Papers' note
The researchers compared methods using synthetic medical records seeded with canaries to detect leakage. Direct fine-tuning delivered the strongest capabilities but leaked the most; DP-SGD reduced leakage at a substantial capability cost. Privacy-aware rewriting protected information better than naive paraphrasing across multiple rewriting models. More capable rewriters generally reduced downstream leakage, but extra test-time compute produced inconsistent gains.
HF Daily Papers' note
score 4