Abstract: Addressing the challenge of dynamically constructing kill-webs for UAV swarms in high-intensity adversarial environments, a collaborative decision-making framework integrating deep reinforcement learning and differential game theory is proposed. This method breaks through the limitations of traditional static optimization, enabling autonomous evolution of the kill-web: an instantaneous decision-making engine transforms battlefield situational awareness into a combinatorial optimization problem, generating periodic greedy policy samples to provide a value baseline for reinforcement learning. The dynamic distributed kill-web allows each UAV to autonomously determine communication nodes and to strike targets based on local situational awareness, supporting global task-oriented topology self-organization and dynamic function self-recovery. An adversarial trajectory closed-loop design provides differentiated motion planning for opposing forces—the offensive side employs a hybrid virtual potential field method for high-speed penetration and evasion, while the defensive side solves optimal cooperative interception trajectories based on differential game theory. This is augmented by a task feedback mechanism connecting the decision-making layer with the motion planning layer. This framework significantly enhances swarm system resilience in highly adversarial and dynamic scenarios, providing a scalable autonomous collaborative solution for intelligent warfare.