Add Gaussian noise calibrated to (ε, δ) to: (a) each empirical-quantile value before serializing the marginal, (b) each Spearman/polyserial estimate before assembling the correlation matrix, (c) each Bernoulli probability. Then project the noisy correlation matrix back to nearest PSD (already done elsewhere). Reference: Frontiers in Digital Health 2025 DP-Gaussian-copula. CLI: syntha fit --epsilon 1.0. Acceptance: at ε=1.0, max KS ≤ 0.05, MIA AUC ≤ 0.55.
Add Gaussian noise calibrated to (ε, δ) to: (a) each empirical-quantile value before serializing the marginal, (b) each Spearman/polyserial estimate before assembling the correlation matrix, (c) each Bernoulli probability. Then project the noisy correlation matrix back to nearest PSD (already done elsewhere). Reference: Frontiers in Digital Health 2025 DP-Gaussian-copula. CLI:
syntha fit --epsilon 1.0. Acceptance: at ε=1.0, max KS ≤ 0.05, MIA AUC ≤ 0.55.