【摘要】 船舶轨迹预测是智能航运的核心技术，但现有模型在处理长序列数据时存在计算开销大、全局时空依赖关系建模能力不足及隐私保护缺失等问题。为此，本文提出基于Mamba-Transformer融合的船舶轨迹预测与差分隐私保护方案。该方案在轨迹预测模块中设计了双路径并行结构，通过Mamba分支的线性扩展能力高效捕捉长程时序依赖，借助Transformer分支强大的全局建模能力提取轨迹的宏观模式，并通过本文设计的分层多头注意力模块实现二者深度融合，进而同时有效捕获轨迹的局部航行细节与全局分布模式。此外，针对实时轨迹预测相比延时发布隐私泄露风险更高的特点，所提方案在模型输出层引入差分隐私保护机制，其设计的基于时间衰减的隐私预算分配策略显著提升了隐私保护下的发布轨迹的效用。基于丹麦海事数据集的实验结果表明，本文所提方案在船舶轨迹预测精度上较现有方法实现了显著提升；同时，通过引入灵活的差分隐私机制，为高精度预测结果提供了严格的隐私保障。更多还原

【Abstract】 Ship trajectory prediction is a core technology for intelligent shipping, but existing models suffered from high computational costs, inadequate modeling of global spatio-temporal dependencies, and lack of privacy protection when processing long-sequence data. To address these challenges, this paper proposed a ship trajectory prediction and privacy-preserving scheme based on Mamba-Transformer fusion. The scheme innovatively designed a dual-path parallel architecture in the trajectory prediction module, efficiently capturing long-range temporal dependencies through the linear scaling capability of the Mamba branch while leveraging the powerful global modeling capability of the Transformer branch to extract macroscopic trajectory patterns. Deep fusion was achieved through a hierarchical multi-head attention module designed, thereby effectively capturing both local navigation details and global trajectory patterns simultaneously. Moreover, to address the higher privacy leakage risks associated with real-time trajectory prediction compared to delayed release, the proposed solution introduced a differential privacy protection mechanism at the model output layer. The designed time-decay-based privacy budget allocation strategy significantly enhanced the utility of published trajectories under privacy protection. Experimental results on the Danish maritime dataset demonstrate that the proposed scheme achieves substantial improvements in ship trajectory prediction accuracy over existing methods while providing rigorous privacy guarantees for high-precision predictions through a flexible differential privacy mechanism.更多还原