Adverse weather conditions significantly degrade the performance of vision-based systems. While recent methods have achieved remarkable results in single-weather removal, their performance drops substantially when handling mixed degradations in real-world scenarios. Although different weather conditions exhibit distinct degradation patterns, they also share underlying common characteristics. To address this issue, we propose a Dynamic Multi-Attention Network (DMANet) to jointly exploit shared knowledge and degradation-specific features. DMANet is built upon dynamic multi-attention blocks (DMABs), which capture multi-dimensional features through adaptive attention mechanisms and dynamic weighting. Specifically, DMAB integrates three complementary attentions: multi-directional attention (MDA) for global structural modeling, local-global attention (LGA) for fine-grained detail enhancement, and frequency-spatial attention (FSA) for dual-domain texture restoration. In addition, a dynamic gating unit (DGU) is introduced to adaptively weight different attention branches according to degradation severity, enabling effective interaction between shared and condition-specific features. An adaptive enhancement module (AEM) further refines feature distributions to balance stability and diversity, thereby improving restoration fidelity. Extensive experiments demonstrate that DMANet achieves state-of-the-art performance for adverse weather removal.