Consensus Computation in Unreliable Networks: A System Theoretic Approach

This work considers the problem of reaching consensus in an unreliable linear consensus network. A solution to this problem is relevant for several tasks in multi-agent systems including motion coordination, clock synchronization, and cooperative estimation. By modeling the unreliable nodes as unknown and unmeasurable inputs affecting the network, we recast the problem into an unknown-input system theoretic framework. Only relying on their direct measurements, the agents detect and identify the misbehaving agents using fault detection and isolation techniques. We consider both the case that misbehaviors are simply caused by faults, or that they are the product of a definite, malignant “Byzantine” strategy. We express the solvability conditions of the two cases in a system theoretic framework, and from a graph theoretic perspective. We show that generically any node can correctly detect and identify the misbehaving agents, provided that the connectivity of the network is sufficiently high. Precisely, for a linear consensus network to be generically resilient to k concurrent faults, the connectivity of the communication graph needs to be 2k + 1, if Byzantine agents are allowed, and k + 1, if non-colluding agents are considered. We finally provide algorithms for detecting and isolating misbehaving agents. The first procedure applies standard fault detection techniques, and affords complete intrusion detection if global knowledge of the graph is available to each agent, at a high computational cost. The second method is designed to exploit the presence in a network of weakly interconnected subparts, and provides computationally efficient detection of misbehaving agents whose behavior deviates more than a threshold, which is quantified in terms of the interconnection structure.