<p>Dynamic hand gesture authentication emerges as a promising contactless biometric method that integrates both physiological and behavioral characteristics. However, due to inter-individual differences in execution speed and key behavior phases, most existing methods rely on fixed-weight temporal aggregation, failing to effectively highlight highly discriminative segments. In addition, conventional global identity aggregation methods often introduce high computational cost and may be insufficient for capturing robust identity representations. To address these challenges, we propose the adaptive temporal weighting network (ATW-Net). The core contribution is the adaptive temporal weighting (ATW) module, which dynamically learns the importance of each temporal segment and enhances the representation of key behavior phases. Meanwhile, the Manhattan Self-Attention (MaSA) module is adopted to achieve efficient global feature aggregation, and C-AdamW is employed to improve training stability. Extensive experiments on the SCUT-DHGA dataset show that ATW-Net achieves an MG EER of 0.518% and a UMG EER of 0.700% with only 11.710M parameters. Ablation experiments, comparative experiments, and visualizations of learned temporal weights further validate the effectiveness of the proposed modules. Experiments on SCUT-DHGA-br also demonstrate the robustness of ATW-Net under complex backgrounds and illumination variations.</p>