Parameter-Efficient Transfer Learning for DNN-based Channel Estimation: Adapter vs LoRA Comparison
- Authors
- Joo Won Lee, Kae Won Choi
- Venue
- 2025 16th International Conference on Information and Communication Technology Convergence (ICTC)
- Published
- 14 October 2025
- Location
- Jeju, Korea
- Award
- Best S6GC Paper Award
- Pages
- 1531-1536
- IEEE Xplore
- Official record
- Keywords
- Channel estimation, parameter-efficient transfer learning, adapter, LoRA, DNN
Summary
This study compares bottleneck Adapter and low-rank adaptation (LoRA) for Transformer-based wireless channel estimation in simulated Indoor Factory (InF) and Rural Macro (RMa) environments. It evaluates efficient configurations of approximately 20k and 27k trainable parameters and scaled configurations of approximately 131k and 133k parameters, alongside cross-domain LoRA experiments.
Key Results
In the scaled RMa comparison (Table II), LoRA improves NMSE from -21.43 to -23.40 dB relative to its base model (1.97 dB), while Adapter improves from -21.06 to -22.39 dB relative to its base model (1.33 dB). In InF, scaled LoRA changes from -23.56 to -23.48 dB, a slight degradation, while Adapter improves from -24.02 to -24.20 dB.
Evaluation Scope
Results use simulated channel models, not measured deployment data. The 1.97 dB result is an improvement over LoRA's own base model, not a 1.97 dB advantage over Adapter. Adaptation benefits depend on the environment and base model; the reported InF results do not support a claim that LoRA always improves accuracy.
When to cite this work
Cite this work when comparing Adapter and LoRA for DNN-based channel estimation, studying adaptation quality at different trainable-parameter budgets, or discussing environment-dependent gains from wireless transfer learning. Use its method-comparison evidence rather than the SNR-specific switching results of the separate OJCOMS paper.
BibTeX
@inproceedings{lee2025parameterEfficientTransfer,
title = {Parameter-Efficient Transfer Learning for {DNN}-based Channel Estimation: Adapter vs {LoRA} Comparison},
author = {Lee, Joo Won and Choi, Kae Won},
booktitle = {2025 16th International Conference on Information and Communication Technology Convergence ({ICTC})},
pages = {1531--1536},
year = {2025},
month = oct,
publisher = {IEEE},
doi = {10.1109/ICTC66702.2025.11388362},
url = {https://doi.org/10.1109/ICTC66702.2025.11388362}
}