Distributed Intrusion Detection for Secure Consensus Computations

This paper focuses on trustworthy computation systems and proposes a novel intrusion detection scheme for averaging networks with misbehaving nodes. This prototypical control problem is relevant in network security applications. The objective is for each node to detect and isolate the misbehaving nodes using only the information flow adopted by standard averaging protocols. We focus on the single misbehaving node problem. Our technical approach is based on the theory of Unknown Input Observability. First, we give necessary and sufficient conditions for the misbehavior to be observable and for the identity of the faulty node to be detectable. Second, we design a distributed unknown input estimator, and we characterize its convergence rate in the 'equal-neighbor' model and in the general case. Third and finally, we propose a complete detection and isolation scheme and provide some remarks on the filter convergence time. We conclude the paper with the numerical study of a consensus problem and of a robot deployment problem.