Cam-PC: A Novel Method for Camouflaging Point Clouds to Counter Adversarial Deception in Remote Sensing

Abstract

Synthetic aperture LiDAR can generate point cloud data, which is widely used in 3-D scene reconstruction. However, existing point cloud object recognition methods are vulnerable to adversarial attacks, and such attacks are difficult to transfer to the physical world. Even if adversarial perturbations are added to physical objects, they are easily detectable by other sensors. Our proposed method includes two modules, R-D and D-R, which generate more concealed adversarial point cloud samples by modifying digital and physical features. The R-D module maps real-world entities to point cloud data in the digital world and generates adversarial samples by modifying signal amplitude values. The D-R module constructs adversarial objects by modifying the surface diffuse reflectance of the target object based on ray tracing and correspondences between digital and physical features. Our method is evaluated through experiments on attack effectiveness, robustness after subsampling and transferability, demonstrating its effectiveness, and achieving new state-of-the-art performance.

Keywords

• Adversarial attack
• physical simulation
• point cloud
• remote sensing image