On Basing ZK ≠ BPP on the Hardness of PAC Learning

Learning is a central task in computer science, and there are various formalisms for capturing the notion. One important model studied in computational learning theory is the PAC model of Valiant (CACM 1984). On the other hand, in cryptography the notion of "learning nothing'' is often modelled by the simulation paradigm: in an interactive protocol, a party learns nothing if it can produce a transcript of the protocol by itself that is indistinguishable from what it gets by interacting with other parties. The most famous example of this paradigm is zero knowledge proofs, introduced by Goldwasser, Micali, and Rackoff (SICOMP 1989). Applebaum et al. (FOCS 2008) observed that a theorem of Ostrovsky and Wigderson (ISTCS 1993) combined with the transformation of one-way functions to pseudo-random functions (Hastad et al. SICOMP 1999, Goldreich et al. J. ACM 1986) implies that if there exist non-trivial languages with zero-knowledge arguments, then no efficient algorithm can PAC learn polynomial-size circuits. They also prove a weak reverse implication, that if a certain non-standard learning task is hard, then zero knowledge is non-trivial. This motivates the question we explore here: can one prove that hardness of PAC learning is equivalent to non-triviality of zero-knowledge? We show that this statement cannot be proven via the following techniques: 1. Relativizing techniques: there exists an oracle relative to which learning polynomial-size circuits is hard and yet the class of languages with zero knowledge arguments is trivial. 2. Semi-black-box techniques: if there is a black-box construction of a zero-knowledge argument for an NP-complete language (possibly with a non-black-box security reduction) based on hardness of PAC learning, then NP has statistical zero knowledge proofs, namely NP is contained in SZK. Under the standard conjecture that NP is not contained in SZK, our results imply that most standard techniques do not suffice to prove the equivalence between the non-triviality of zero knowledge and the hardness of PAC learning. Our results hold even when considering non-uniform hardness of PAC learning with membership queries. In addition, our technique relies on a new kind of separating oracle that may be of independent interest.