Optimal Byzantine Attack for Distributed Inference with M-ary Quantized Data

Abstract

In many applications that employ wireless sensor networks (WSNs), robustness of distributed inference against Byzantine attacks is important. In this work, distributed inference is considered when local sensors send M-ary data to the fusion center. The optimal Byzantine attack policy is then derived under the assumption that the Byzantine adversary has the knowledge of the statistics of local quantization outputs. Our analysis indicates that the fusion center can be blinded such that the detection error is as poor as a random guess when an adequate fraction of sensors are compromised.