Agile Semantics Alignment over Fading Channel via LoRA-based Fine-Tuning

초록

This paper proposes a communication and learning framework to rapidly resolve semantics misalignment (SMA) between a TX and an RX in dynamic fading channels. The SMA arises from a mismatch in feature vector dimensions when wireless channel quality degrades. Our approach is based on a split-architecture semantic communication system, where the transmitter (TX) uses principal component analysis (PCA) for dimensionality reduction. The receiver (RX) then reconstructs the transmitted feature vector and performs fine-tuning on its pre-trained model using Low-Rank Adaptation (LoRA) to swiftly adapt to the changing channel conditions. Experimental results on a V2X testbed show that our method achieves significant latency reduction compared to end-to-end training, while maintaining stable and high accuracy compared to using a pre-trained model without fine-tuning. © 2025 IEEE.

키워드