Consensus Breakdown for Countering Malicious Large-Scale MASs Via Graph Attention Network

Multi-agent systems (MASs) are increasingly exploited for malicious purposes, posing significant threats to public safety and national security. Existing countermeasures, which are effective for single or small groups of agents, often fail to deal with large-scale MASs due to the complexity and scale of their network topologies. To address this challenge, this paper proposes a rapid adversarial approach aimed at decomposing the network topology of MASs by disabling the critical nodes, thereby preventing MASs from achieving global consensus. This approach transforms the critical node identification problem into a node regression task and introduces the Multi-Head Graph Attention Network (Multi-head GAT) for feature learning of the network topology, which is particularly well-suited for capturing the complex dependencies and diverse relationships within large-scale networks. The multi-head attention mechanism of Multi-head GAT allows the model to capture information at different scales, enhancing the proposed algorithm’s robustness and accuracy. Based on the strength of the Multi-head GAT, the importance scores of agent nodes are efficiently calculated by the proposed algorithm to enable precise and effective network topology decomposition. Extensive simulations demonstrate that our approach significantly disrupts MASs’ coordination capabilities by preventing global consensus with superior performance across various scenarios. Comparative experiments further validate the accuracy and efficiency of the proposed method.