Communication Locality in Secure Multi-party Computation - How to Run Sublinear Algorithms in a Distributed Setting

We devise multi-party computation protocols for general secure function evaluation with the property that each party is only required to communicate with a small number of dynamically chosen parties. More explicitly, starting with n parties connected via a complete and synchronous network, our protocol requires each party to send messages to (and process messages from) at most polylog(n) other parties using polylog(n) rounds. It achieves secure computation of any polynomial-time computable randomized function f under cryptographic assumptions, and tolerates up to $({1\over 3} - \epsilon) \cdot n$ statically scheduled Byzantine faults. We then focus on the particularly interesting setting in which the function to be computed is a sublinear algorithm: An evaluation of f depends on the inputs of at most q=o(n) of the parties, where the identity of these parties can be chosen randomly and possibly adaptively. Typically, q=polylog(n). While the sublinear query complexity of f makes it possible in principle to dramatically reduce the communication complexity of our general protocol, the challenge is to achieve this while maintaining security: in particular, while keeping the identities of the selected inputs completely hidden. We solve this challenge, and we provide a protocol for securely computing such sublinear f that runs in polylog(n)+O(q) rounds, has each party communicating with at most q ·polylog(n) other parties, and supports message sizespolylog(n) ·(l+n), where l is the parties' input size. Our optimized protocols rely on a multi-signature scheme, fully homomorphic encryption (FHE), and simulation-sound adaptive NIZK arguments. However, we remark that multi-signatures and FHE are used to obtain our bounds on message size and round complexity. Assuming only standard digital signatures and public-key encryption, one can still obtain the property that each party only communicates with polylog(n) other parties. We emphasize that the scheduling of faults can depend on the initial PKI setup of digital signatures and the NIZK parameters.