5
$\begingroup$

I am interested in the setting of differential privacy- let's say a random function $\mathcal{D}:X\to\mathbb{R}$ discriminates between (distinct) $x, y \in X$ in a differentially private way if $$ \mathbb{P}(\mathcal{D}(x) \in S) \le e^\epsilon \mathbb{P}(\mathcal{D}(y) \in S) $$ for all (Borel) subsets $S$ and also the reverse inequality holds ($\epsilon>0$ is fixed). (The prototypical example is $\mathcal{D}(x)=f(x) + $Laplace noise but I don't want the question to focus on that.)

This definition implies an upper bound on the most powerful test for $H_0:x=y$ vs. $H_A: x \neq y$ in terms of $\epsilon$ (see later in question for a naive bound). I am not familiar with the literature and would appreciate a reference for the best known bound.

In a concrete setting with Laplace noise I have estimated the max power below. For this question I am interested in general bounds with mild assumptions. On the other hand, the more work I look at this, the more I doubt much better general bounds are possible!


Edit [naive bound]: For the purposes of obtaining the bound N-P applies- after all we are just testing the distribution of $\mathcal{D}(x)$ against $\mathcal{D}(y)$.

Let's say I just test whether a single observation $o\in S$ is distributed according to $X$ ($H_0$) or $Y$ ($H_A$) and assume discreteness, for convenience of presentation. This is a N-P setting so the likelihood ratio gives the best test. $$ L(o):=\frac{\mathbb{P}(X=o)}{\mathbb{P}(Y=o)}. $$

Let the critical (rejection) region at $\alpha$ be $C$, i.e. $$ \alpha = \mathbb{P}(L(X) \in C) = \mathbb{P}(X \in L^{-1}(C)). $$ Now apply the DP definition with $S=L^{-1}(C)$ to get $$ \mathbb{P}(L(Y) \in C) = \mathbb{P}(Y \in L^{-1}(C)) \le \alpha e^\epsilon. $$

The bound derived this way for $n$ observations is $$ \mathbb{P}(L(Y_1,\ldots,Y_n) \in C) \le \alpha e^{n \epsilon}, $$
using hopefully obvious notation.

But for $n$ and $\epsilon$s I have seen real life $\alpha e^{n\epsilon} > 1$ so the bound is useless. However, I can't see how to easily improve it. I don't mind some extra assumptions on $\mathcal{D}$, e.g. regularity or discreteness.


Edit2 [Comparing naive bound to actual best for $n=1$ Laplace noise] I checked the $n=1$ test explicitly with Laplacian noise (assuming parameters are known and the median differs by exactly 1, a situation that satisfies DP with scale=$1/\epsilon$). When $\epsilon$ is small (i.e. the Laplace variance is large) the log likelihood ratio is basically delta at $\pm\epsilon$ and a critical region for $\alpha < e^{-2\epsilon}/2$ does not exist. When $\epsilon$ is large the Laplace distribution is highly concentrated around its median and the critical value is around $e^{-\epsilon}$ ~$0$ and test power is ~$.5$. For moderate $\epsilon$ the bound from naive application of the DP definition is good.

So in practice the best test power looks like below ($\alpha=0.05$ and we accept using a bigger critical region for small $\epsilon$):

power for Laplace noise test


Edit3 [Naive bound is awful as $n$ increases] As the number of observations increases the naive bound becomes lousier and lousier. By $n=10$ observations it is already useless. Below power is estimated by simulation and plotted against empirical alpha$\times exp(n\epsilon)$ because an exact $0.05$ critical region doesn't exist for small alpha and epsilon.

Naive bound lousy for $n>1$


Edit4 [Estimate max power for Laplace noise using CLT approximation to log LR]

For small - moderate $\epsilon$ (s.t. the Laplace noise is reasonably flat in a neighbourhood of the median) I approximated each term in the log likelihood ratio under the null hypothesis by $$ \ln(L_i) = \begin{cases} +\epsilon & \text{w.p.} 1/2 \\ -\epsilon & \text{w.p.} e^{-\epsilon}/2 \\ U[-\epsilon,\epsilon] & \text{otherwise} \\ \end{cases} $$ and going from null to alternative hypothesis amounts to simply switching the sign (using symmetry of the Laplace distribution). This just a sum of nice distributions and the central limit theorem has already kicked in strongly around $n=30$, i.e. the null distribution is approximately Gaussian with mean $n\epsilon (1 - e^{-\epsilon})/2$ and variance given by a hideous formula. The simulation below is with $n=30$.

enter image description here

Using this I can estimate the power for larger $n$ (plot below). This basically answers the question for the Laplace mechanism, but I'm still interested in what can be deduced from the DP definition.

enter image description here

$\endgroup$
2
1
$\begingroup$

If I understand your question correctly, you're looking for a characterization of differential privacy in the formalism of hypothesis testing. This idea was formalized by Kairouz et al. in The Composition Theorem for Differential Privacy; see Section 2 and Theorem 2.1 in particular.

$\endgroup$
2
  • $\begingroup$ Thanks Ted for the reply. I am not working on this now but skimmed the paper out of curiosity. Test power is $1 - $type II error, i.e. $1 - P_{MD}$ in the paper so my question could be rephrased "what is the type II error of the best test?", which Theorem 2.1 doesn't seem to answer - rather it's more like "differential privacy is equivalent to preventing both low type I and II error", and only tells me $P_{MD}\geq e^{-\epsilon}(1 - P_{FA})$ or power $\leq 1 - e^{-\epsilon}(1 - P_{FA})$. If I can stomach a high type I error this doesn't tell me much (and for my setting that was OK) $\endgroup$ – P.Windridge Apr 25 '20 at 9:44
  • $\begingroup$ Hope I didn't miss anything and +1 for the reference :) $\endgroup$ – P.Windridge Apr 25 '20 at 9:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.