UAM Cybersecurity Dataset

A Cross-Layer CNS Attack Dataset for Passenger eVTOL Operations

High-Fidelity eVTOL Simulation · Communication–Navigation–Surveillance Attacks · Log-Verifiable Labels · Seoul Urban Low-Altitude Corridor

Download Dataset — Coming Soon

1. Overview

This dataset captures cyberattacks on a passenger eVTOL (BETA Alia-250) cruising along a low-altitude urban corridor over Seoul, from Seongsan Bridge to Yeongdong Bridge. It was collected in a co-simulation environment of X-Plane 12, PX4 SITL, and QGroundControl. Wind sampled from 10 years of Seoul winter wind statistics was applied to every flight. A comprehensive set of flight-state variables was logged throughout each flight.

It contains five attacks spanning all three CNS layers (Communication, Navigation, and Surveillance). All five were collected on the same aircraft, route, and environmental distribution. Each attack is grounded in real aviation incidents or attacks demonstrated on real avionics hardware.

To the best of our knowledge, this is the first public dataset that combines the UAM operational domain with the CNS attack surface.

2. Key Features

Cross-Layer CNS Coverage

Attacks target the C2 link, GNSS, VOR/DME, and onboard altitude state. Prior public datasets cover a single layer, and none include surveillance-layer attacks.

Incident-Grounded Attacks

Attacks are based on the 2016 Melbourne spoofed ATC instructions, the 2023 GPS spoofing near Baghdad, the 2024–2025 Incheon FIR GNSS interference, and demonstrated avionics data-bus altitude falsification.

High State Observability

A broad set of simulator state variables is recorded, covering the vehicle's internal state in detail. PX4 estimated state is logged alongside the simulator's true state.

Log-Verifiable Labels

The manipulated variable of each attack is recorded in the log, so every labeled interval can be verified directly from the data. Users can also redefine labels with other variables.

Realistic Variability

Wind is sampled from measured Seoul wind distributions. Attacks repeat within a flight, producing multiple normal → attack → recovery transitions.

Route-Leakage-Aware Evaluation

All flights follow the same route, so a model could learn route position instead of vehicle state. Evaluation code that prevents this shortcut is included.

3. Dataset Composition

Label Classes

Configuration	Scenario	Flights
PX4	Normal	1,800
PX4	GPS Spoofing	200
PX4	GPS Jamming	200
PX4	Command Injection	200
X-Plane	Normal	1,800
X-Plane	COM/NAV Disruption	200
X-Plane	Altitude Manipulation	200
Total		4,600

Each attack is compared only against normal flights from its own configuration.

4. Attack Scenarios

Attack	CNS Layer	Method	Observed Response
Command Injection	Communication	Forces a fixed-wing → hover mode transition during cruise, then resumes the mission. Repeated twice per flight	Mission halt, abnormal loitering in the corridor
GPS Spoofing	Navigation	Gradual position bias in HIL_GPS lat/lon (≈40 s attack / ≈60 s pause, 3 cycles), including slow drift of hundreds of meters	Gradual path deviation, hard to distinguish from normal
GPS Jamming	Navigation	Full or partial degradation of GNSS fix quality, 6 events per flight (≈10 s each)	Suppressed lateral correction, then a correction burst after jamming ends
COM/NAV Disruption	Navigation	Injects an out-of-band NAV2 frequency (400.00 MHz) into standby and flips it to active	DME drops to 0; VOR bearing freezes
Altitude Manipulation	Surveillance	Falsified indicated altitude (1,200 ft, then 800 ft; planned 1,000 ft)	Autopilot moves the aircraft opposite to the falsified value (≈243–349 m)

GPS Spoofing and GPS Jamming assume an external RF attacker. The other three assume an internal attacker with access to onboard data paths. All attacks begin after the aircraft enters stable cruise.

5. Ground-Truth Labeling

Labels are derived from the manipulated variable recorded in the log: modified HIL_GPS fields and injected commands in the PX4 configuration, and modified DataRefs in the X-Plane configuration.

For GPS Jamming, the post-jamming interval carries a separate Recovery label, because the vehicle's response peaks after jamming ends.

6. Download & Publication

Dataset Download: Coming Soon

Publication: Coming Soon

모빌리티 사이버보안 연구실

DATASET