Consider a joint distribution $(X,A)$ on a set ${\cal X}\times\{0,1\}^\ell$. We show that for any family ${\cal F}$ of distinguishers $f \colon {\cal X} \times \{0,1\}^\ell \rightarrow \{0,1\}$, there exists a simulator $h \colon {\cal X} \rightarrow \{0,1\}^\ell$ such that
\begin{enumerate}
\item no function in ${\cal F}$ can distinguish $(X,A)$ from $(X,h(X))$ with advantage $\epsilon$,
\item $h$ is only $O(2^{3\ell}\epsilon^{-2})$ times less efficient than the functions in ${\cal F}$.
\end{enumerate}
For the most interesting settings of the parameters (in particular, the cryptographic case where $X$ has superlogarithmic min-entropy, $\epsilon > 0$ is negligible and ${\cal F}$ consists of circuits of polynomial size), we can make the simulator $h$ \emph{deterministic}.
As an illustrative application of this theorem, we give a new security proof for the leakage-resilient stream-cipher from Eurocrypt'09. Our proof is simpler and quantitatively much better than the original proof using the dense model theorem, giving meaningful security guarantees if instantiated with a standard blockcipher like AES.
Subsequent to this work, Chung, Lui and Pass gave an interactive variant of our main theorem, and used it to investigate weak notions of Zero-Knowledge. Vadhan and Zheng give a more constructive version of our theorem using their new uniform min-max theorem.
↧