Paper 2025/817

Relating Definitions of Computational Differential Privacy in Wider Parameter Regimes

Abstract

The literature on computational differential privacy (CDP) has focused almost exclusively on definitions that are computational analogs of `pure' $(\epsilon,0)$-DP. We initiate the formal study of computational versions of approximate DP, i.e. $(\epsilon, \delta)$-DP with non-negligible $\delta$. We focus on IND-CDP and SIM$_{\forall\exists}$-CDP and show that the hierarchy between them when $\delta > 0$ potentially differs substantially from when $\delta = 0$. In one direction, we show that for $\delta < 1$, any mechanism which is $(\epsilon,\delta)$-SIM$_{\forall\exists}$-CDP also is $(\epsilon,\delta)$-IND-CDP, but only if $\epsilon$ is logarithmic in the security parameter. As a special case, this proves that the existing implication from $(\epsilon,0)$-SIM$_{\forall\exists}$-CDP to $(\epsilon,0)$-IND-CDP does not hold for arbitrary $\epsilon$, as previously claimed. Furthermore, we prove that when the parameters are the same in IND-CDP and SIM$_{\forall\exists}$-CDP and $\epsilon$ is superlogarithmic, there exists a natural task that can be solved whilst satisfying SIM$_{\forall\exists}$-CDP but which no IND-CDP mechanism can solve. This is the first separation in the CDP literature which is not due to using a task contrived specifically in order to give rise to the separation. In the other direction, we show that the techniques for establishing an implication from $(\epsilon,0)$-IND-CDP to $(\epsilon,0)$-SIM$_{\forall\exists}$-CDP extend only to that a mechanism being $(\epsilon,\delta)$-IND-CDP implies it is also $(\epsilon,\delta')$-SIM$_{\forall\exists}$-CDP with $\delta' > \delta$. Finally, we show that the Groce-Katz-Yerukhimovich barrier results against separations between CDP and statistical DP hold also in the setting of non-negligible $\delta$.