Subspace Differential Privacy

Many data applications have certain invariant constraints due to practical needs. Data curators who employ differential privacy need respect such on the sanitized product as a primary utility requirement. Invariants challenge formulation, implementation, and interpretation of guarantees. We propose subspace privacy, honestly characterize dependence output confidential aspects data. discuss two design frameworks that convert well-known differentially private mechanisms, Gaussian Laplace ones invariants specified by curator. For linear queries, we near-optimal mechanisms minimize mean squared error. Subspace rid for post-processing invariants, preserve transparency statistical intelligibility output, can be suitable distributed implementation. showcase proposed 2020 Census Disclosure Avoidance demonstration data, spatio-temporal dataset mobile access point connections large university campus.