MARS

MARS: Defending Unmanned Aerial Vehicles From Attacks on Inertial Sensors with Model-based Anomaly Detection and Recovery

Overview

We introduce a general attack-resilient state estimator that does not require inertial sensors (e.g., IMUs) to estimate UAV system states, and incorporates position and heading sensors to generate high-fidelity body-frame orientation and angular velocity estimates when inertia sensors fail.
We design a MARS-based CUSUM sliding window intrusion/anomaly detector (ID) that outperforms existing anomaly detectors by swiftly and accurately detecting IMU attacks and providing system-level alarms for UAV security-aware decision-making.
We develop a multi-stage flight recovery strategy that prioritizes security without significantly sacrificing performance. Our method is universally applicable to arbitrary UAV missions, regardless of the UAV's state, ensuring freedom in UAV maneuverability when equipped with the MARS detection and recovery framework.
We evaluate MARS methodology on the PX4 open-source autopilot firmware in large-scale simulations as well as real-world physical experiments, we demonstrate MARS effectiveness, even against attacks for which other security methods fail.

Open Science Policy

To comply with the open science policy, we will ensure that our source code and datasets for both simulations and real-world experiments are publicly available on GitHub by the deadline specified in the submission policy.

Simulation

Comparison with state-of-art defense methods.

Hovering Stabilization under Attack

hover_lpf_crashed.MP4

LPF

A simple and widely used frequency-based filter. Crashed.

hover_dae_crashed.MP4

DAE

A powerful machine learning-based filter. Crashed.

hover_caf_crashed.MP4

CAF

A complimentary attitude feedback approach using geometric relationship between attitude and position measurements. Crashed.

hover_ours.MP4

MARS (Ours)

Our proposed method using rotor speed sensors and an attack-resilient state estimator. Survived.

In this analysis, we deployed them in PX4 autopilot and tested them in SITL hovering missions with attacks. The quadcopter is set to take off and hover at a given location when IMU attacks are initiated.
The figure tracks the altitude of the quadcopter in the hovering mission under the four recovery frameworks when four attack vectors are injected. The three representative methods failed to maintain the drone's stability, causing it to fall to the ground within 2 senconds, while MARS made the UAV survive the attacks.

2. Dynamic Mission Recovery

tracking_ours.MP4

In this analysis, the quadcopter was tasked with completing a tracking mission along a 10 m trajectory on the X-axis in the Earth frame while maintaining a fixed altitude. Attacks on the inertial sensors were initiated mid-mission, and our objective is to complete the mission without crashing.
MARS successfully implemented the multi-stage dynamic recovery process, which included braking when the attacks were detected, restoring to a near-hovering condition, and then continuing the mission in a recovered flight.

Real-world Case Study

Hovering Mission

hovering.mp4

Hovering

In this case, an intermittent IMU saturation attack at 1.5s period with 0.5s attack pulses is injected. It shows MARS’s capability in consecutive detection and recovery periods.

Monitor-Left: MARS anomaly detector status with attack flag (attacker perspective), anomaly detection rate and anomaly flag (system perspective).
Monitor-Middle: PX4 IMU-based standard angular velocity estimate error. It represents the estimate quality in the performance-prioritized standard control mode and it easily falls victim to attacks.
Monitor-Right: MARS angular velocity estimate error. It represents the estimate quality in security-prioritized resilient control. With suboptimal performance, It is robust to attacks on IMU.

2. Dynamical Missions

waypoint.mp4

Waypoint Visiting

To demonstrate the effectiveness of MARS dynamical recovery, we launch attacks when the drone is perform a waypoint visiting mission. The attack profile is AR-DoS. Upon detection of the attack, the drone is swiftly switched to hovering to minimize oscillation caused by the attack. After clearing the attack, the drone resumes its mission.

Monitor-1: MARS anomaly detector status.
Monitor-2: PX4 IMU-based standard angular velocity estimate error.
Monitor-3: MARS angular velocity estimate error.
Monitor-4: Drone 3d real-time position.

square.mp4

Square Tracking

Next, we evaluate MARS dynamic recovery in a more challenging scenario, the square tracking. Four 0.5-second AR-DoS attacks are injected during the mission while the drone manages to complete the task with minimal performance loss.

Monitor-1: MARS anomaly detector status.
Monitor-2: PX4 IMU-based standard angular velocity estimate error.
Monitor-3: MARS angular velocity estimate error.
Monitor-4: Drone 3d real-time position.

3. Hovering Mission under Wind Disturbance

hovering_wind.mp4

Wind Disturbance Test

To further demonstrate MARS's robustness to dynamical changing environments, we created wind disturbance using a wind blower.
The wind blower was set directly pointing at the drone from a distance of 3m, at the same height of UAV flying, and produced a wind disturbance at around 3m/s at the hovering position, measured by a handheld anemometer.
This mission shows that despite the increment of position deviation and angular tilts due to the wind, MARS managed to control the drone with acceptable stability.

4. Outdoor Tests

outdoor_hovering_with_gps.mp4

Outdoor Test 1: Hovering

To test MARS under a more realistic settings, we enabled the onboard GPS module on our MARS-PX4 quadcopter and conducted an outdoor experiment.
This mission demonstrates that MARS is still effective after switching from indoor lab settings to outdoor field test conditions.
Monitor-1: MARS anomaly detector.
Monitor-2: Standard angular velocity estimate (affected by attacks).
Monitor-3: MARS angular velocity estimate.
Monitor-4: The GPS 2D position in the local frame, transformed from global latitude and longitude (for anonymity purposes). The red dashed circle represents the maximum distance the drone drifts during hovering.

outdoor_waypoint_visiting_with_subtitles.mp4

Outdoor Test 2:
Waypoint Visiting

The same attack strategy is applied as the one carried out indoors.
This test demonstrates MARS's dynamic recovery capability under the more challenging outdoor conditions. The recovery process is executed effectively, remaining unaffected by changing environmental factors. It shows MARS dynamic recovery framework is a robust real-world solution.

outdoor_square_tracking_with_subtitles.mp4

Outdoor Test 3:
Square Tracking

The same attack strategy is applied as the one carried out indoors.
This test showcases MARS's anomaly detection and recovery performance in a more challenging outdoor tracking mission. The attacks are initiated while the quadcopter is maneuvering in different directions. Nevertheless, MARS successfully detects all attacks and maintains stable flight with minimal performance degradation.

Page updated

Google Sites

Report abuse