Abstract: Objectives To address the challenge of time-varying and uncertain intrusion behavior in highly dynamic and strongly adversarial environments, this paper proposes a method for joint intention and trajectory prediction based on deep learning. Methods The adversarial scenario is modeled as a dynamic interaction graph sequence comprising multiple entities such as defending vessels, intruding vessels, obstacles, and protected areas, with four types of interaction edges defined: confrontation, cooperation, penetration, and evasion. A spatio-temporal graph convolutional network is employed to capture short-term strategic interaction features, while a bidirectional long short-term memory network is introduced to model the temporal evolution of behaviors. An intention-guided dual-branch decoder is designed to achieve collaborative output and joint optimization of intention recognition and trajectory prediction. Training and validation are conducted using both simulated data and harbor prototype vessel data. Results Simulation results show that the proposed method achieves an intention recognition accuracy of 94.9%, outperforming the baseline algorithm (90.3%). Over a 10-step prediction horizon, the ratio of average displacement error to travel distance (RADE) is 18.7%. In the harbor prototype vessel experiment, the intention recognition accuracy reaches 88%, with an RADE of 21.6%, and the performance advantage over the baseline remains stable.Conclusions The proposed method effectively leverages adversarial interaction information, enhances situational awareness in complex adversarial environments, and provides support for collaborative decision-making.