Non-Colluding Attacks Identification in Distributed Computing

Abstract

This paper studies a distributed computing setting in which the computing task consists of multiplying a matrix by a vector. A number of worker machines are attacked, i.e., the result of their computation is maliciously perturbed by some adversaries. In particular, the focus is on the case where these adversaries are non-colluding and non-communicating and hence they cannot jointly perturb the results of all the attacked worker machines. First, a condition that ensures that the result of the computing task can be successfully recovered with high probability is derived as a function of the setting parameters. Then, a probabilistic mechanism inspired by group testing is proposed to identify the set of the attacked worker machines, and the corresponding probability of error is derived.