We show that indistinguishability obfuscation implies that
all functions with sufficient ``pseudo-entropy'' cannot be obfuscated
under a virtual black box definition with a universal simulator. Let
${\cal F}=\{f_s\}$ be a circuit family with super-polynomial
pseudo-entropy, and suppose ${\cal O}$ is a candidate obfuscator with
universal simulator $\Sim$. We demonstrate the existence of an adversary $\Adv$ that, given the obfuscation ${\cal O}(f_s)$, learns a predicate the simulator $\Sim$ cannot learn from the code of $\Adv$ and black-box access to $f_s$. Furthermore, this is true in a strong sense: for \emph{any} secret predicate $P$ that is not learnable from black-box access to $f_s$, there exists an adversary that given ${\cal O}(f_s)$ efficiently recovers $P(s)$, whereas given oracle access to $f_s$ and given the code of the adversary, it is computationally hard to recover $P(s)$.
We obtain this result by exploiting a connection between obfuscation with a universal simulator and obfuscation with auxiliary inputs, and by showing new impossibility results for obfuscation with auxiliary inputs.
↧