Attack-Resilient Multi-Agent Flocking Control Using Graph Neural Networks

Abstract

Flocking control of a group of mobile agents has been recently investigated using Graph Convolution Networks (GCNs). The design relies on training using a centralized controller but the resulting GCN controller is based on communication between the agents. The agents receive sensor measurements which are incorporated into the states and shared between the neighbors. However, the paradigm is prone to adversarial attacks. In this paper, we consider the problem of designing GCN-based distributed flocking control that is resilient to attacks on the communicated information. We consider an attack model that is used to compromise the inter-agent communication and may inject arbitrary signals. Our control design uses a coordinate-wise median-based aggregation function. It is shown that the GCN-based controller using the proposed aggregation method is resilient against attacks on the communication between the agents, whereas the typical average-based aggregation fails to maintain the flock structure. Robustness analysis is performed to show that the proposed method is resilient whenever a majority of the agents in the neighborhood can be trusted. Simulation results and analysis are presented that validate the merits of the proposed approach.