In recent years, multi-modal recognition has attracted extensive attention due to its security and high accuracy. However, existing dual-modal fusion methods are limited, and simple fusion strategies fail to meet the requirements of recognition tasks. To address this issue, this paper proposes an Attention-guided Multi-level Fusion Module (AFM), which integrates both shallow and deep features from two modalities through collaborative learning rather than simple concatenation or addition. This fusion approach maximizes the complementary characteristics of fingerprint and finger-vein information, enhancing the discriminative power of the fused features. Additionally, we introduce the RSANet with attention mechanisms. Experimental results demonstrate that RSANet achieves 98.4% identification accuracy on the SDUMLA dataset, 100% on the FVC-HKPU dataset, and 99.5% on the NUPT-FPV dataset. These results validate the effectiveness of the proposed method. Ablation studies on the AFM module further confirm that the multi-level fusion significantly improves recognition performance.