Solanki, Abhijeet; Amiri, Wesam Al; Hasan, Syed Rafay; Guo, Terry N. (2026).听.听IEEE Access.听
Autonomous vehicles (AVs) use multiple sensors鈥攅specially聽cameras听补苍诲听LiDAR聽(a laser-based system that measures distance)鈥攖o understand their surroundings. However, camera systems can be tricked by聽laser-based attacks, where directed light creates misleading patterns in images. Earlier versions of these attacks required constant adjustment, making them difficult to use in real time. This study introduces a simpler and more practical method: a聽static laser attackthat does not need continuous tuning but can still disrupt how the vehicle detects objects.
The researchers analyzed how these laser attacks affect images by examining their visual patterns using statistical measures, image-processing techniques, and machine learning methods. They found that even fixed laser interference can reduce object recognition accuracy and interfere with how camera and LiDAR data are combined (sensor fusion). Some attacks create obvious artifacts that can be detected using basic image filters and color-based analysis, while more subtle ones require more advanced, geometry-aware detection methods.
To address both types, the authors developed a lightweight detection system that combines statistical and geometric features to identify these attacks in real time. Their approach achieved over 90% detection accuracy on a compact computing device, showing that it is both effective and practical for use in real-world autonomous systems.
FIGURE 1.听
Static attack visualization. Each row shows a normal clean image (top) and its corresponding attacked version (bottom). Static AdvLB introduces consistent bright streaks, while Static AdvLS adds reproducible neon-green circular patterns. These fixed-parameter attacks are real-time deployable and highly reproducible.