Feature Sparsification Based on Feature Importance for Wireless Split Learning

Split Learning (SL) has emerged as a promising paradigm for distributed model training in resource-constrained Internet-of-Things (IoT) environments by partitioning deep neural networks between lightweight client devices and a powerful central server. However, SL suffers from substantial communication overhead due to the frequent transmission of high-dimensional intermediate activations. In this paper, we propose a novel sparsification framework based on Gradient-weighted Class Activation Mapping (Grad-CAM) to alleviate this bottleneck. By leveraging the Grad-CAM, our approach quantifies the importance of activations and selectively transmits only the most discriminative intermediate activations. To overcome the limitation that clients lack direct access to Grad-CAM scores during the current forward pass, we utilize the channel importance vectors computed at the server from the previous iteration as surrogates, thereby enabling effective sparsification without incurring additional computational overhead. Extensive experiments demonstrate that our method significantly reduces communication cost compared to baseline schemes. Furthermore, our results reveal a trade-off between mini-batch size and sparsification ratio in SL, emphasizing the importance of careful activation selection for robust learning under communication constraints.