Trajectory Prediction of Dynamic UAV Swarm with Interaction and Quantity Uncertainty under Saturation Attack Mission

Saturation attack missions (SAM) represent a typical paradigm in modern unmanned aerial vehicle (UAV) confrontations. Considering the heterogeneity of dynamic inter-agent interactions and the uncertainty of group size during adversarial processes, accurate trajectory prediction for such dynamic swarms remains an unresolved challenge, particularly in scenarios where the interaction relationships are not directly observable. In this paper, we propose a Masked Dynamic Heterogeneous Interaction Modeling with Attention-based Spatio-Temporal Message Passing (MDHIM). It incorporates dynamic heterogeneous edge embedding aggregation to robustly infer variable interaction relationships from historical trajectories. An attention-enhanced spatio-temporal message passing mechanism is designed to reduce cumulative errors during long-term multi-step prediction. In addition, a state masking strategy is applied to handle dynamically varying swarm sizes caused by agent attrition. Comprehensive experiments conducted on a specially constructed SAM UAV trajectory dataset demonstrate that MDHIM significantly outperforms state-of-the-art baselines across several key metrics for interaction inference, trajectory prediction, and task-level effectiveness. This work provides a robust solution for trajectory prediction in dynamic and adversarial UAV swarms.