Topological Attention Graph Neural ODE Deep Clustering for UAV Swarms in Encirclement Attack Scenarios

In encirclement attacks scenarios, hostile unmanned aerial vehicles swarms commonly employ coordinated tactics of decoy and strike formations. This tendency poses significant challenges to traditional defense systems with insufficient group identification precision. Deep clustering algorithms have gradually emerged as a critical approach to solving such problem due to their advantages in complex feature disentanglement. Existing methods typically utilize independent encoders to learn node features and employ shallow graph neural networks (GNNs) to extract topological features. Because attributes and graph structure are tightly coupled in many scenarios, this separation suppresses structure-aware attribute propagation and results in biased embeddings. Moreover, the shallow architectures of traditional GNNs fundamentally limit their capacity to extract global collaborative features from multihop neighbors. To address these issues, this article proposes a topological attention graph ordinary differential equation (ODE) deep clustering network. Specifically, we introduce an ODE-enhanced topological attention mechanism to capture both local and global features of swarm. Furthermore, we couple the extraction processes of node features and topological features through an autoencoder architecture. Finally, we define a spectral clustering loss function that reduces the influence of initial cluster centers by exploiting properties of the similarity matrix. Experiments on encirclement attack datasets demonstrate that our algorithm outperforms State-of-the-Art baselines in performance metrics.