UAV-CAS

UAV-CAS (UAV Collaborative Attack Scenarios) is a large-scale labeled flow dataset for intrusion detection in UAV swarm networks, generated by a Containernet digital twin that is systematically calibrated against measurements from the NSF AERPAW testbed.

Why it exists

IDS models trained on wired-network benchmarks such as CICIDS2017 and UNSW-NB15 degrade sharply in real UAV swarms, where mobility, fluctuating link quality, and decentralized routing reshape traffic distributions. Existing UAV-specific datasets do not systematically vary those conditions — which leaves no way to train or test a detector against the very shift that defeats it. They also do not cover collaborative attacks, where several adversaries coordinate different attack types whose combination is invisible to a single-flow detector.

Collecting this on physical hardware is not practical: it needs sustained flight hours, dedicated spectrum and airspace, and the ability to safely execute attacks that a physical testbed will not permit. UAV-CAS closes the gap in simulation and then argues the fidelity case with measurements rather than assertion.

What’s in it

The configuration grid varies four axes independently, so shift can be attributed rather than merely observed:

• swarm size — 5 vs. 20 drones (the dominant source of distributional diversity)
• transmit power — 10 vs. 30 dBm
• mission profile — spiral vs. random mobility
• path-loss model — log-distance vs. 3GPP

No single axis accounts for the dataset’s diversity, which rules out the dataset’s spread being an artifact of one over-weighted parameter.

Calibration — the fidelity argument

A four-layer pipeline validates the twin against AERPAW at increasing levels of abstraction: altitude-dependent path loss, mission-specific mobility, the link-level performance chain, and end-to-end trace fidelity.

Simulated-to-real RSS divergence measures 0.33 Hellinger. The reference point that makes that number interpretable is the 0.73 divergence between two independent real measurement campaigns — so the gap between the twin and the hardware is smaller than the gap between two hardware measurement runs.

Benchmark: what 10 baseline IDS reveal

Binary attack detection is a learnability floor, not the operational task. On binary benign-vs-DoS, nine of ten architectures exceed 0.99 weighted F1 and the lowest reaches 0.989 — which confirms the data is clean and learnable, and also explains why the field can look solved.

The difficulty is in the full multi-class taxonomy. Per-class F1 ranges from near zero to 0.82 and falls into the single digits for stealth attacks. Stealth-paired collaborative compositions — blackhole+wormhole, blackhole+replay — collapse to 0.51–0.86 AUROC.

The design intent is visible in the statistics: standard deviations on flow-level features are wider for stealth attacks (blackhole, wormhole) than for floods, because stealth attacks are built to mimic the natural variability of benign UAV traffic. High-rate DoS and DDoS flows form compact clusters in a t-SNE projection; stealth flows do not. When both components of a composition are individually near-benign, no single-flow feature suffices — detection has to reach for cross-flow or temporal-context signal.

Intended use

• training and evaluating IDS under controlled distribution shift, with shift attributable to a known configuration axis
• collaborative / multi-adversary detection, including set-recovery formulations where the label is the composition rather than a single class
• early-detection studies from short packet prefixes
• stress-testing detectors that report strong binary numbers on legacy benchmarks
• transfer and generalization studies between wired, IoT, and aerial network traffic

Citation

@article{mishra2026uavcas,
  title   = {UAV-CAS: A Calibrated Digital-Twin Dataset for Intrusion
             Detection in UAV Swarm Networks},
  author  = {Mishra, Sripath and Bhargava, Bharat and Liu, Zizheng
             and Islam, Shafkat},
  journal = {arXiv preprint arXiv:2606.17845},
  year    = {2026}
}

Dataset DOI: 10.21227/zgrg-z865