Abstract: The increasing size and computational complexity of deep neural networks (DNNs) impose substantial demands on computing resources, storage capacity, and energy consumption, which restrict their efficient deployment on resource-constrained edge devices. To address the challenges of maintaining inference accuracy during model compression and adapting models to heterogeneous hardware resources, an attention-based dynamic feature pruning and adaptive deployment method is proposed. A channel-spatial attention module is first constructed to dynamically evaluate feature importance from both channel and spatial dimensions, enabling redundant features to be selectively pruned while preserving informative representations and reducing computational overhead. Furthermore, an adaptive deployment framework (ADF) is designed to generate multiple submodels with different compression configurations through a single training process. According to the available computing and storage resources of target edge devices, an appropriate submodel can be selected for deployment, thereby improving hardware adaptability. Experimental results on the CIFAR-10 dataset show that the proposed method reduces floating-point operations (FLOPs) by approximately 48% while causing only a 0.2% decrease in accuracy. These results demonstrate that combining attention-guided dynamic feature pruning with adaptive deployment can effectively reduce model computational complexity while maintaining performance, and can improve the deployment efficiency and flexibility of DNNs in heterogeneous edge computing environments.