Beyond the Glow: Understanding Luminescent Marker Behavior Against Autonomous...
summary
Paper demonstrates that simple luminescent lane paints and projected UV/laser illumination create non‑uniform glow and splat pattern textures that break modern lane detectors: digital tests produce up to complete model failure (worst cases) and about 33% failure in best cases, TwinLightNet fails frequently, YOLOPV2 degrades, ClearerNet is inconsistent. Physical tests on a device running Openpilot reproduce misdetections that cause off‑lane or braking hazards. Proposed mitigations include adversarial training and synthetic augmentation with diffusion/generative models.
AI-written summary · about
chapters (9)
- 0:00Luminescent markers and AV perception
- 1:17Lane detection and modern models
- 3:25Luminescent markers for testing
- 4:30Threat model and attack scenarios
- 6:38Digital simulation setup
- 8:02Physical testing approach
- 8:45Evaluation metrics and performance
- 10:08Physical setup results and implications
- 10:50Conclusion and future research
related talks
14mSeeing is Deceiving: Systematic Vulnerability Analysis of LiDAR-Based Autonomous...
10mInvestigating Stealthy Latency Attacks Against Camera-Based Object...
01h 06mSelf-Driving and Connected Cars: Fooling Sensors and Tracking Drivers
55mPerception Deception: Physical Adversarial Attack Challenges
46mML Use Cases at Think Big Analytics
51mTeaching Autonomous Vehicles to Drive with Visible and Non-Visible Light Simulation
59mStudying Failure Modes of Deep Learning and Ways To Fix Them
06mWhen Collaboration Betrays: The Evolving Attack Surface of...