Multimodal biometric recognition has been widely applied in various fields in recent years due to its advantages in enhancing the security and convenience of identity verification. However, traditional multimodal systems face the challenge of information redundancy. Existing methods rely on static teacher models for self-distillation and neglect the quality differences in inter- and intra- modalities, which hinder their overall performance. To address these issues, we propose a dynamic selective distillation network (DSD-Network) that integrates the early exit mechanism with dynamic selective self-distillation to optimize the performance and computational efficiency of multimodal biometric recognition frameworks. Specifically, we introduce a selective distillation strategy based on entropy-based to regulate the application of distillation. Subsequently, a dynamic distillation factor is employed to adaptively adjust the distillation process according to the learning status of each layer. The integration enables early-stage predictions by transferring knowledge from deeper layers to shallow layers, thereby further enhancing the model’s performance and efficiency. Additionally, a Modality Quality-Aware Fusion (MQAF) approach is introduced, which incorporates intra-modality confidence and inter-modality energy uncertainty weighting. This module leverages an inter-modality energy uncertainty evaluation strategy to calibrate intra-modality feature weights, fully exploiting the complementarity and correlation among modalities. Extensive experiments have been conducted to validate the effectiveness of the proposed method.