On Malicious Agents in Non-Bayesian Social Learning with Uncertain Models

Abstract

Social Learning is the process of cooperatively aggregating information between agents in order to collectively estimate or learn an unknown value. Most all research in social learning assume that the likelihoods of the private observations given the possible hypotheses are known with absolute certainty. However, these likelihoods must be machine learned before the social learning process. Recent work has extended social learning for uncertain likelihoods. Such likelihoods are only known within Dirichlet distributions due to limited training samples available to learn them. This paper investigates the effects of malicious agents when both the good and malicious agents are uncertain about their likelihoods. Such malicious agents are trying to drive the consensus to accept an incorrect hypothesis and reject the correct hypothesis. This paper also presents and evaluates a method to identify and remediate against the effects of the malicious agents.