A new preprint on arXiv tackles a core problem in differential privacy: when randomized response perturbs data to protect individuals, it also distorts the statistical relationships among variables. The authors derive an exact attenuation formula for multivariate cross-moments under this privacy mechanism, giving analysts a precise way to understand how much signal is lost.
The work centers on an alternative formulation of randomization matrices that remains compliant with differential privacy. By recasting the perturbation process, the authors are able to characterize how joint moments shrink or shift, rather than relying on approximations or empirical corrections.
This is a technical contribution aimed at improving the utility of privatized analyses. If the formula holds in practice, it could let researchers recover closer estimates of covariances and higher-order relationships from data that has been deliberately obscured.