Byzantine Multiple Access

TL;DR

This paper investigates reliable and authenticated communication over multiple access channels with potential adversarial users, characterizing capacity regions and proposing optimal schemes for both scenarios.

Contribution

It provides the first capacity region characterization for randomized and deterministic codes in adversarial MACs, and introduces a rate-optimal authentication scheme with minimal overhead.

Findings

Capacity region characterized for randomized codes.

Deterministic codes achieve capacity when non-symmetrizable.

Authenticated capacity is either zero or equals the MAC capacity.

Abstract

We study communication over multiple access channels (MAC) where one of the users is possibly adversarial. When all users are non-adversarial, we want their messages to be decoded reliably. When an adversary is present, we consider two different decoding guarantees. In part I, we require that the honest users' messages be decoded reliably. We study the 3-user MAC; 2-user MAC capacity follows from point-to-point AVC capacity. We characterize the capacity region for randomized codes. We also study the capacity region for deterministic codes. We obtain necessary conditions including a new non-symmetrizability condition for the capacity region to be non-trivial. We show that when none of the users are symmetrizable, the randomized coding capacity region is also achievable with deterministic codes. In part II, we consider the weaker goal of authenticated communication where we only require that an adversarial user must not be able to cause an undetected error on the honest users' messages. For the 2-user MAC, we show that the following 3-phase scheme is rate-optimal: a standard MAC code is first used to achieve unauthenticated communication followed by two authentication phases where each user authenticates their message treating the other user as a possible adversary. We show that the authentication phases can be very short since this form of authentication itself, when possible, can be achieved for message sets whose size grow doubly exponentially in blocklength. This leads to our result that the authenticated communication capacity region of a discrete memoryless MAC is either zero or the (unauthenticated) MAC capacity region itself. This also, arguably, explains the similar nature of authenticated communication capacity of a discrete memoryless point-to-point adversarial channel recently found by Kosut and Kliewer (ITW, 2018).