Blind Spot in Sensor Fusion: Exploiting an Architectural Vulnerability to Hijack...
summary
Shows that PX4 and ArduPilot’s architecture leaves a persistent blind spot: injected gyroscope bias immediately alters control outputs while GPS-based fusion lags, enabling sustained steering. Builds a black-box reinforcement-learning attack using an asymmetric actor-critic that modulates gyroscope bias (bounded ±0.06 rad/s bias every second) to steer a hovering UAV 200 m toward a 10 m-radius target, achieving 85–87% success in simulation, bypassing detectors and succeeding 7/7 on real hardware.
AI-written summary · about
chapters (10)
- 0:00Sensor fusion vulnerability research
- 1:14UAV steering attack requirements
- 2:40Gyroscope correction delay
- 4:46Gyroscope attack threat model
- 6:10GAP with reinforcement learning
- 7:55Performance vs baseline attacks
- 9:43GAP robustness and transferability
- 10:26GAP vs existing defenses
- 11:12UAV flight demo results
- 11:57Summary and conclusion
related talks
09mWIP: Hijacking Attacks on UAV Follow-Me Systems in Realistic Scenarios
17mConfuSense: Sensor Reconfiguration Attacks for Stealthy UAV Manipulation
11mWIP: Evaluating the End-to-End Impact of False Localization Attacks on vSLAM-...
53mEfficient Visual Avoidance and Control for UAVs with Embedded GPUs
01h 06mSelf-Driving and Connected Cars: Fooling Sensors and Tracking Drivers
21mInjected and Delivered: Fabricating Implicit Control over Actuation Systems by Spoofing Inertial Sensors
26mLegal GNSS Spoofing and its Effects on Autonomous Vehicles
40mSecuring Tesla and Mobileye from Split-Second Phantom Attacks