Local open‐ and closed‐loop manipulation of multiagent networks

Abstract

The manipulation of networked cyberphysical devices via local external actuation or feedback control is explored, in the context of a canonical multiagent dynamical system that is engaged in a consensus or synchronization task. One main focus is to understand whether or not, and how easily, a stakeholder can manipulate the full network's dynamics by hijacking only one agent's actuation signal. Explicit spectral characterizations are given of the energy (effort) required to manipulate the dynamics. These characterizations are used to (1) gain structural insights into ease of manipulation, (2) show that manipulation along the consensus manifold is easy, and (3) address network design to enable or prevent manipulation. Additionally, it is shown that the multiagent system can be manipulated effectively along the consensus manifold using local feedback controls, which do not require model knowledge or wide-area measurements.SAHABANDU ET AL. topology among the network's components. Thus, local actions in a CPN incur patterned wide-area impacts, regardless of the specifics of the local action or the network dynamics. In this work, we seek to understand whether and how local actuations can be designed to manipulate the network's wide-area behavior and to evaluate extreme impacts (whether beneficial or harmful). To do this, we consider a linear double-integrator network (DIN) model, which serves as a common abstraction for a number of CPN processes defined on a graph (eg, vehicle traffic on a highway, formation-controlled aerial vehicle teams, linearized swing dynamics of the electric power grid, and certain sensor-network algorithms). [6][7][8] This article explores the manipulation of the DIN via external actuation or feedback control at a single agent. First, we study whether or not the actuation signal for an agent can be designed to move the network's state to an arbitrary value and also the actuation effort (energy) required for manipulation. Characterizations are obtained in terms of the network's coupling parameters and the location of the manipulated agent. As a comparison, manipulation via a local feedback control is also considered, wherein only local measurements are used to decide the additional actuation of the agent. This is done by analyzing the transfer function of the dynamics from the perspective of the control channel, which then indicates whether or not the system can be manipulated to achieve certain control objectives (eg, fast tracking or disturbance rejection) in a feedback setting.The research described here contributes to an important research thrust on the security of cyberphysical systems (CPSs), which is a key property required for their effective operation (eg, the work of Banerjee et al 9 ). CPS security encompasses security of information, ie, the protection of sensitive or private information that is held by the CPS. In complement, CPS security also encompasses functional security, ie, protection of CPS functions from attacks and failures. The need for functional security for CPS is pa...

Controllability-Gramian Submatrices for a Network Consensus Model

“…Because of the relevance of Gramian matrices to network controllability as well as dual observability/estimation problems, some structural and graph-theoretic results on the full Gramian have been developed for canonical network-consensus models, as well as for other dynamical network processes [9,11,16,17,20].…”