Causality-inspired kinematics-constrained adversarial trajectory generation for swarm intention recognition

Deep neural networks increasingly support Swarm Intention Recognition (SIR) from multi-agent trajectories, yet their adversarial robustness is difficult to assess because position, velocity, and acceleration (PVA) are temporally coupled. Coordinate-wise perturbations can disrupt these relationships and yield trajectories with limited physical meaning. We propose Causality-Inspired Trajectory Adversarial Generation (CTAG), a kinematics-constrained white-box framework that optimizes a bounded upstream motion variable and reconstructs downstream states through differentiable integration. CTAG uses acceleration for PVA inputs and velocity for position–velocity inputs. Sequential projection enforces the intervention bound, and an optional iterative minimum-separation repair supports explicit constraint handling. A multi-dimensional protocol relates attack success to displacement, topology, rollout consistency, detectability, scalability, and closed-loop replay behavior. The evaluation covers three swarm-trajectory families and five time-series backbones. CTAG achieves macro attack success rates of 0.980 on UavMix3T20 and 0.965 on AeroSwarmN50T20, within 0.020 and 0.035 of the strongest direct-state results under clean-centered full-input constraints. It yields the smallest absolute change in position rollout-consistency gap among eight primary methods, reaching 0.260 m and 0.229 m on the two datasets. Closed-loop 6-DOF replay yields a realized attack success rate of 0.881 and retains 89.7% of successful reference attacks, with no divergence, saturation, collision, or separation violation in the evaluated CTAG cohort. Together, these results support high attack effectiveness, lower position rollout-consistency changes, and simulator-scoped closed-loop realization under the evaluated conditions.