Generalized Universal Circuits for Secure Evaluation of Private Functions with Application to Data Classification

Abstract

Secure Evaluation of Private Functions (PF-SFE) allows two parties to compute a private function which is known by one party only on private data of both. It is known that PF-SFE can be reduced to Secure Function Evaluation (SFE) of a Universal Circuit (UC). Previous UC constructions only simulated circuits with gates of d = 2 inputs while gates with d > 2 inputs were decomposed into many gates with 2 inputs which is inefficient for large d as the size of UC heavily depends on the number of gates.We present generalized UC constructions to efficiently simulate any circuit with gates of d ≥ 2 inputs having efficient circuit representation. Our constructions are non-trivial generalizations of previously known UC constructions.As application we show how to securely evaluate private functions such as neural networks (NN) which are increasingly used in commercial applications. Our provably secure PF-SFE protocol needs only one round in the semi-honest model (or even no online communication at all using non-interactive oblivious transfer) and evaluates a generalized UC that entirely hides the structure of the private NN. This enables applications like privacy-preserving data classification based on private NNs without trusted third party while simultaneously protecting user’s data and NN owner’s intellectual property.Keywordsuniversal circuitssecure evaluation of private functionsneural networksprivate data classificationprivacy