Compressing Unit-Vector Correlations via Sparse Pseudorandom Generators

A unit-vector (UV) correlation is an additive secret-sharing of a vector of length B that contains 1 in a secret random position and 0's elsewhere. UV correlations are a useful resource for many cryptographic applications, including low-communication secure multiparty computation and multi-server private information retrieval. However, current practical methods for securely generating UV correlations involve a significant communication cost per instance, and become even more expensive when requiring security against malicious parties.

In this work, we present a new approach for constructing a pseudorandom correlation generator (PCG) for securely generating n independent instances of UV correlations of any polynomial length B. Such a PCG compresses the n UV instances into correlated seeds whose length is sublinear in the description size n • log B. Our new PCGs apply in both the honest-majority and dishonest-majority settings, and are based on a variety of assumptions. In particular, in the honest-majority case they only require "unstructured" assumptions. Our PCGs give rise to secure endto-end protocols for generating n instances of UV correlations with o(n) bits of communication. This applies even to an authenticated variant of UV correlations, which is useful for security against malicious parties. Unlike previous theoretical solutions, some instances of our PCGs offer good concrete efficiency.

Our technical approach is based on combining a low-degree sparse pseudorandom generator, mapping a sparse seed to a pseudorandom sparse output, with homomorphic secret sharing for low-degree polynomials. We then reduce such sparse PRGs to local PRGs over large alphabets, and explore old and new approaches for maximizing the stretch of such PRGs while minimizing their locality.

Finally, towards further compressing the PCG seeds, we present a new PRG-based construction of a multiparty distributed point function (DPF), whose outputs are degree-1 Shamir-shares of a secret point function. This result is independently motivated by other DPF applications.