Identifying Reliable Machines for Distributed Matrix-Vector Multiplication

This paper considers a distributed computing framework, where the task of T matrix-vector products is distributed among n worker machines. External adversaries have access to a subset ℒ (the cardinality of which is |ℒ|) of these machines, and can maliciously perturb the result of each of their computations with probability α. To correctly recover each matrixvector product, the master has to identify a set (of a fixed cardinality) of ‘unattacked’ worker machines. Towards this end, this work proposes four schemes that aim at performing such an identification. These schemes are analyzed and compared under different regimes of (|ℒ|,α) for the two cases when |ℒ| is (1) known or (2) unknown at the master.