kangnguyen/LoRaSeek_SF7_Large

LoRaSeek: Boosting Denoising Ability in Neural-enhanced LoRa Decoder via Hierarchical Feature Extraction

ACM MobiCom 2025

Khang Nguyen¹, Yidong Ren¹, Jialuo Du¹, Jingkai Lin¹
Maolin Gan¹, Shigang Chen², Mi Zhang³, Chunyi Peng⁴, Zhichao Cao¹

¹ Michigan State University · ² University of Florida · ³ The Ohio State University · ⁴ Purdue University

LoRaSeek uses hierarchical feature extraction to improve the denoising and representation capability of neural LoRa decoders, particularly at low signal-to-noise ratios.

Overview

• A hierarchical U-Net with CNN and hybrid Transformer for multi-scale signal representation
• A hybrid, lightweight Transformer with channel scaling and local-enhanced FFN.
• Dual attention-based skip connections for preserving important chirp characteristics across scales
• Decoding using 2 options: LoRaPHY vs. Light-weight DNN Citation

Model Details

• Model: LoRaSeek
• Variant: Large
• Checkpoint: SF7_Large
• LoRa Spreading Factor: SF7
• Framework: PyTorch
• Task: LoRa/CSS symbol demodulation
• Signal bandwidth: 125 kHz
• Input representation: Time-frequency representation derived from received LoRa signals

If you use this model or LoRaSeek in your research, please cite the corresponding LoRaSeek paper:

@inproceedings{nguyen2025loraseek,
  title={LoRaSeek: Boosting denoising ability in neural-enhanced LoRa decoder via hierarchical feature extraction},
  author={Nguyen, Khang and Ren, Yidong and Du, Jialuo and Lin, Jingkai and Gan, Maolin and Chen, Shigang and Zhang, Mi and Peng, Chunyi and Cao, Zhichao},
  booktitle={Proceedings of the 31st Annual International Conference on Mobile Computing and Networking},
  pages={712--726},
  year={2025}
}