Official experiment code and confirmatory result artifacts for UCS-RFNet: Reliability-Aware Device-to-Device RF Collaboration for Low-Altitude Unauthorized UAV Detection.
UCS-RFNet combines directly supervised local RF classifiers, uncertainty-aware sparse feature selection, link-aware graph fusion, and exact packet serialization for collaborative UAV detection over a flying ad hoc network (FANET). The repository preserves the validation-first protocol used for the manuscript and publishes the aggregate evidence needed to reproduce its tables and figures.
All values below are mean ± sample standard deviation over five prespecified seeds.
| Method | Detection | FAR | Macro-F1 | Payload (KB/interval) | Reported time (ms) |
|---|---|---|---|---|---|
| Single Tiny-CNN | 0.8188 ± 0.0112 | 0.0494 ± 0.0025 | 0.8264 ± 0.0076 | 0.0000 | 0.717 ± 0.009 |
| Full feature sharing | 0.8491 ± 0.0149 | 0.0505 ± 0.0021 | 0.8431 ± 0.0099 | 1.1016 | 11.007 ± 0.075 |
| Magnitude Top-K | 0.8480 ± 0.0136 | 0.0505 ± 0.0024 | 0.8428 ± 0.0092 | 0.6016 | 6.928 ± 0.055 |
| UCS-RFNet | 0.8499 ± 0.0151 | 0.0510 ± 0.0018 | 0.8425 ± 0.0087 | 0.6016 | 6.943 ± 0.066 |
Relative to Single Tiny-CNN, UCS-RFNet gains 1.61 percentage points in Macro-F1 (paired 95% CI: 1.18 to 2.05 pp). It satisfies the declared −0.5 pp non-inferiority margins against the validation-selected Magnitude Top-K reference, while reducing byte-counted payload by 45.4% relative to dense feature sharing. These are the claims supported by the locked protocol; the results do not establish superiority over every low-payload baseline.
“Reported time” is simulator wall time on one GPU plus modeled airtime for a 1 Mbit/s shared broadcast channel. It is not a physical multi-device end-to-end latency measurement.
UCS-RFNet/
├── src/ucs_rfnet/ # Models, data pipeline, training, metrics, audits
├── configs/ # Frozen confirmatory and validation-candidate YAML
├── scripts/ # Windows/Linux experiment entry points
├── tests/ # Unit and smoke tests
├── analysis/build_paper_assets.py # Rebuilds paper tables, figures, and evidence
├── artifacts/round0_confirmatory_v3/
│ ├── tables/ # Curated aggregate/run-level CSV files
│ ├── validation/ # Validation-only selection evidence
│ ├── ACCEPTANCE_GATE.json # Prespecified claim gate (PASS)
│ ├── INTEGRITY_AUDIT.json # Independent metric/integrity audit (pass)
│ └── PROTOCOL_LOCK.json # Frozen protocol digest
├── paper_assets/generated/ # Regenerated PDF/PNG figures and LaTeX rows
└── docs/ # Reproduction and evidence documentation
The public repository intentionally excludes the third-party RF dataset, checkpoints, receiver-level predictions, caches, and machine logs. Their expected classes and counts are documented in docs/EXPERIMENT_ARTIFACT_MANIFEST.md. Aggregate and seed-level tables required for the manuscript analysis are included.
Python 3.11 is recommended.
git clone https://github.com/Dr-zhaoyf/UCS-RFNet.git
cd UCS-RFNet
python -m venv .venv
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"
pytest -qTo validate the published aggregate bundle and regenerate the manuscript figures and LaTeX rows:
python analysis/build_paper_assets.pyThe analysis script checks the acceptance and integrity verdicts, exact row counts, seed coverage, finite values, published means, and paired confidence intervals before writing any derived asset.
The confirmatory run uses the public Noisy Drone RF Signal Classification dataset. The dataset is not redistributed here; obtain it under its upstream terms.
python -m ucs_rfnet.download_data --source kaggle_noisy_v1 --data-root ../dataThe frozen configuration expects the repository and data/ to be sibling directories. The complete validation → protocol lock → smoke test → one-time confirmatory test → supplement workflow is documented in docs/REPRODUCIBILITY.md. A detailed Chinese execution guide is also available in docs/REMOTE_EXPERIMENT_GUIDE.zh-CN.md.
- Claim gate:
PASS; maximum seed FAR 0.05360 under the 0.055 cap. - Integrity audit:
pass; 1,035 metrics independently recomputed from receiver-level predictions with maximum absolute discrepancy1.11e-16. - Data: 98,705 real labeled RF samples; receiver/channel/FANET diversity is simulated.
- Split: deterministic class-stratified sample-index split. The upstream tensor does not provide session, recording, or physical-receiver group identifiers.
- Generalization: channel, topology, view-count, and node/link perturbations are covered; held-out-device generalization is not established.
See docs/RESULTS.md for the statistical interpretation and docs/RESULTS_EVIDENCE.md for the generated evidence ledger.
Please use CITATION.cff. Bibliographic venue and DOI information can be added when the manuscript is published.
The code is released under the MIT License. Dataset files remain subject to their upstream license and terms.