Mixture Mechanisms Improve Gaussian Differential Privacy

Researchers have introduced a new type of additive noise method called mixture mechanisms, which improves utility for (ε, δ)-differential privacy (DP) in scalar real-valued queries with known sensitivity. By blending several Gaussian distributions that have the same variance but different means and mixture weights, these mechanisms outperform the traditional analytic Gaussian method, especially in settings with moderate to low privacy. These distributions can be seen as a mix of a zero-mean Gaussian and others based on query sensitivity. The authors also set strict variance criteria for (ε, δ)-DP and provide effective algorithms for calculation. Mixture mechanisms notably decrease expected noise levels and variances compared to the analytic Gaussian option. This study is accessible on arXiv with the reference 2605.28078.