Securing Tesla and Mobileye from Split-Second Phantom Attacks
summary
Researchers demonstrate split-second phantom attacks that remotely project images (drone or hacked digital billboard) to trigger real-world ADAS: Mobileye 630 reads a projected speed sign as real at 125 ms detection threshold, Tesla Model X (HW2.5) treats projected person/car as real and can brake or stop in the road. They build a lightweight four-network verifier called Ghostbuster (≈1M parameters) that checks context, surface, light and depth; it reduces baseline attack success from 92–99% to below 1% while keeping false alarms negligible.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:52Perception of Computerized Driver Intelligence
- 4:17Open Virtual Perception Challenges
- 5:08Open Challenges for Virtual Perception
- 5:24Adversarial Attacks on Object Detectors
- 6:02Disadvantages of Adversarial Attacks
- 8:23Stages of Object Detection
- 10:07Persistency Function
- 12:19Significance of Split-Second Phantom Attacks
- 13:00Why Does a Commercial ADAS detect Phantom Objects?
- 15:56Why Are Phantoms a Scientific Bug?
- 18:28Detecting the Appearance Threshold
- 26:03The Different Perspectives
related talks
01h 06mSelf-Driving and Connected Cars: Fooling Sensors and Tracking Drivers
14mSeeing is Deceiving: Systematic Vulnerability Analysis of LiDAR-Based Autonomous...
16mPerspective-Shift Attacks Against Optical Perception Sensors: A Novel Attack...
55mPerception Deception: Physical Adversarial Attack Challenges
43mConnected Cars - What could possibly go wrong?
17mMIRAGE: Detecting Fake Emergency Electronic Brake Light Attacks in V2X Networks...
09mWIP: Hijacking Attacks on UAV Follow-Me Systems in Realistic Scenarios
45mAnalogy Cyber Security—From 0101 to Mixed Signals