ST-SFLora cuts edge AI fine-tuning costs with semantic token selection

Deploying large Transformer-based vision models on resource-limited mobile edge devices remains a major challenge due to hardware constraints and dynamic wireless environments. Federated learning (FL) allows collaborative training without sharing raw data, but local fine-tuning of massive models is computationally prohibitive for edge devices. Split federated learning (SFL) offloads deep layers to an edge server, yet suffers from heavy communication overhead when transmitting high-dimensional activation tokens.

To address this, the authors introduce ST-SFLora, a semantic token-based split federated LoRA fine-tuning framework. They propose a new metric called Semantic Transmission Efficiency (STE) to balance semantic retention and transmission cost. Based on STE, they formulate a joint resource optimization problem that dynamically selects tokens, allocates uplink bandwidth, and sets transmit power under strict latency and energy constraints. The resulting mixed-integer nonconvex problem is solved with an alternating algorithm. Experiments show ST-SFLora achieves the lowest client-side resource consumption among baselines while delivering a favorable trade-off between communication efficiency and model performance.