This paper investigates the cooperative active defense guidance problem in an “evader-pursuer-defender” multi-role game scenario. In practical applications, the hypersonic vehicle and active defender face significant challenges due to uncertain, incomplete, discontinuous and delayed (UID2) information, complicating the evader’s escape. To tackle these issues, a cooperative active defense guidance based on a value decomposition multi-agent reinforcement learning (MARL) algorithm is proposed. First, a reward function is designed specifically for the multi-role cooperative mission, aligning with vehicle maneuverability and game relationships. This helps improve algorithm convergence during training in dynamic and adversarial game scenarios and overcome the limitations of sparse and potentially deceptive reward functions. Second, to handle the spatiotemporal characteristics of the UID2 information, an information processing mechanism is developed to transform detection data into a multi-agent observation set, providing robust support for decision-making. Additionally, an improved value decomposition Qmix-based network architecture is designed to facilitate cooperative strategies during training and enable independent decision-making, guiding the evader and defender in coordinated maneuvers. Finally, the effectiveness of the proposed method is validated through numerical simulations.