Our approach is to generate a synthetic dataset that approximates many randomly-chosen marginal distributions of the input dataset.