FuseFi: Combining Irregularly Sampled CSI from Diverse Communication Packets and Frequency Bands for Wi-Fi Sensing

Reference implementation of the time-aware DNN from "UniFi: Combining Irregularly Sampled CSI from Diverse Communication Packets and Frequency Bands for Wi-Fi Sensing" — Gaofeng Dong, Kang Yang, Mani Srivastava (ECE, UCLA).

Paper: https://arxiv.org/abs/2512.22143

Overview

Existing Wi-Fi sensing systems inject high-rate probing packets (100–1000 Hz) to obtain uniformly sampled Channel State Information (CSI), costing 40%+ of communication throughput. FuseFi instead learns directly from the irregularly sampled CSI that already exists in normal communication traffic — across data, management and control frames on both the 2.4 and 5 GHz bands — with no packet injection.

FuseFi has two stages:

1. CSI sanitization (offline, MATLAB): clusters packets by PHY format, normalizes amplitude, aligns waveforms, removes bursty packets and optionally selects subcarriers. Its output is the .mat dataset consumed here.
2. Time-aware DNN (this repo, PyTorch): learns from non-uniform timestamps without resampling. It fuses CSI and time encodings into content-aware attention keys and drops the missingness mask as an input feature (CSI missingness is not motion-related). A GRU + MLP head produces the activity prediction.

Install

conda env create -f environment.yml && conda activate FuseFi
# or, into an existing environment:
pip install -r requirements.txt

Tested with Python 3.12.1, PyTorch 2.2.1, CUDA 11.8.

Usage

cd src
python tanenc_classification.py

Override only what you want to change:

# a quick smoke test: one seed, a few epochs
python tanenc_classification.py --niters 5 --seed-len 1

Repository layout

src/tanenc_classification.py   train/eval loop (entry point)
src/models.py                  enc_mtan_classif_csi = FuseFi DNN,
                               multiTimeAttention, CombineEmbed
src/utils.py                   dataloaders, burst filtering / union-timeline
                               batching, normalization, evaluate_classifier
src/testbed_dataset.py         Testbed_Dataset_IrregularNSS (CommCSI-HAR reader)

Citation

@article{dong2025unifi,
  title   = {UniFi: Combining Irregularly Sampled CSI from Diverse Communication
             Packets and Frequency Bands for Wi-Fi Sensing},
  author  = {Dong, Gaofeng and Yang, Kang and Srivastava, Mani},
  journal = {arXiv preprint arXiv:2512.22143},
  year    = {2025}
}

Acknowledgements

Built on mTAN (Shukla & Marlin, ICLR 2021).