Noise reduction for Pufferfish privacy: a practical noise calibration method

A practical noise calibration method improves data utility while satisfying Pufferfish privacy by relaxing the overly strict conditions in the 1-Wasserstein (Kantorovich) mechanism that cause excessive noise. The method proves that for all privacy budgets ε and prior beliefs, noise reduction relative to the original 1-Wasserstein mechanism always exists, and the gain is more significant in low privacy budget scenarios; the worst-case 1-Wasserstein mechanism is equivalent to the ℓ1-sensitivity method. On three real datasets, data utility improves by 47% to 87%.