Learning and resilience to information deception in adversarial networked systems: the necessity and structure of a trusted core

We investigate disturbed inference and learning problems in networked systems with adversaries. We Analyze the effects of adversarial attacks on the solutions and characterize the solution robustness and resiliency as functions of network topology and adversary distribution. We demonstrate that existence of a small subnetwork of “trusted nodes” (trusted core) could provide substantial improvements to solution robustness and resilience. We characterize these improvements as functions of the degree of trust, connectivities and location of trusted nodes.