Arithmetic Network Coding for Secret Sum Computation

TL;DR

This paper studies secure sum computation in network coding, proposing a conjecture for Gaussian signals and proving it in cases with up to two wiretapped edges, relevant for sensor networks and federated learning.

Contribution

It introduces a conjecture on secure sum computation for Gaussian signals and proves it for scenarios with at most two wiretapped edges, advancing understanding of secrecy in network coding.

Findings

Conjecture on necessary and sufficient conditions for Gaussian case.

Proof of the conjecture when wiretapped edges are at most two.

Relevance to sensor networks and federated learning.

Abstract

We consider a network coding problem where the destination wants to recover the sum of the signals (Gaussian random variables or random finite field elements) at all the source nodes, but the sum must be kept secret from an eavesdropper that can wiretap on a subset of edges. This setting arises naturally in sensor networks and federated learning, where the secrecy of the sum of the signals (e.g. weights, gradients) may be desired. While the case for finite field can be solved, the case for Gaussian random variables is surprisingly difficult. We give a simple conjecture on the necessary and sufficient condition under which such secret computation is possible for the Gaussian case, and prove the conjecture when the number of wiretapped edges is at most 2.