Witness Encryption and NP-Hardness of Learning

We study connections between two fundamental questions from computer science theory. (1) Is witness encryption possible for NP [Sanjam Garg et al., 2013]? That is, given an instance x of an NP-complete language L, can one encrypt a secret message with security contingent on the ability to provide a witness for x ∈ L? (2) Is computational learning (in the sense of [Leslie G. Valiant, 1984; Michael J. Kearns et al., 1994]) hard for NP? That is, is there a polynomial-time reduction from instances of L to instances of learning? Our main contribution is that certain formulations of NP-hardness of learning characterize the existence of witness encryption for NP. More specifically, we show: - witness encryption for a language L ∈ NP is equivalent to a half-Levin reduction from L to the Computational Gap Learning problem (denoted CGL [Benny Applebaum et al., 2008]), where a half-Levin reduction is the same as a Levin reduction but only required to preserve witnesses in one direction, and CGL formalizes agnostic learning as a decision problem. We show versions of the statement above for witness encryption secure against non-uniform and uniform adversaries. We also show that witness encryption for NP with ciphertexts of logarithmic length, along with a circuit lower bound for E, are together equivalent to NP-hardness of a generalized promise version of MCSP. We complement the above with a number of unconditional NP-hardness results for agnostic PAC learning. Extending a result of [Shuichi Hirahara, 2022] to the standard setting of boolean circuits, we show NP-hardness of "semi-proper" learning. Namely: - for some polynomial s, it is NP-hard to agnostically learn circuits of size s(n) by circuits of size s(n)⋅ n^{1/(log log n)^O(1)}. Looking beyond the computational model of standard boolean circuits enables us to prove NP-hardness of improper learning (ie. without a restriction on the size of hypothesis returned by the learner). We obtain such results for: - learning circuits with oracle access to a given randomly sampled string, and - learning RAM programs. In particular, we show that a variant of MINLT [Ker-I Ko, 1991] for RAM programs is NP-hard with parameters corresponding to the setting of improper learning. We view these results as partial progress toward the ultimate goal of showing NP-hardness of learning boolean circuits in an improper setting. Lastly, we give some consequences of NP-hardness of learning for private- and public-key cryptography. Improving a main result of [Benny Applebaum et al., 2008], we show that if improper agnostic PAC learning is NP-hard under a randomized non-adaptive reduction (with some restrictions), then NP ⊈ BPP implies the existence of i.o. one-way functions. In contrast, if CGL is NP-hard under a half-Levin reduction, then NP ⊈ BPP implies the existence of i.o. public-key encryption.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.