LCC-SL: Learning-Centric Compression for Communication-Efficient Split Learning

Abstract

The rapid proliferation of Internet of Things (IoT) devices is accelerating the development of distributed machine learning paradigms, such as split learning (SL). However, as the number of participating devices increases, SL faces severe communication challenges due to the high-dimensional smashed data transmitted over bandwidth-limited wireless networks. Existing compression methods mainly treat smashed data as generic numerical tensors and ignore task-relevant semantic information, leading to critical activation features to be discarded under aggressive compression. To address this issue, we propose LCC-SL, a learning-centric compression framework for SL, which comprises two key components: semantic activation compression (SAC) module and error-feedback gradient compression (EFGC) module. SAC employs a lightweight semantic encoder–decoder to project high-dimensional activations into compact latent representations, which are then vector-quantized and transmitted as discrete code indices instead of dense activation tensors. EFGC targets backward communication by compressing gradients with residual error feedback, thereby reducing gradient transmission while preserving accumulated optimization information. Extensive experiments validate that LCC-SL effectively reduces the communication overhead while maintaining competitive model accuracy.