Communication-Efficient Semantic Split Federated Learning via Variational Information Bottleneck Driven Dynamic Slicing

Abstract

This letter proposes a novel channel-aware semantic SFL (CA-SSFL) framework. The client model employs a dual-masking architecture, where a distributed variational information bottleneck (VIB) generates a semantic mask by prioritizing feature dimensions, while a signal-to-noise ratio (SNR)-driven module produces a dynamic channel mask reflecting the transmission budget. The element-wise product of the semantic mask and channel mask enables a dynamic slicing mechanism that suppresses irrelevant or unreliable features under current channel conditions. Extensive experiments on the CIFAR-10 and CIFAR-100 datasets under non-IID settings demonstrate that the proposed CA-SSFL drastically reduces communication costs while achieving a favorable accuracy-communication trade-off under severe noise conditions.