A Physically Feasible Counter-Attack Method for Remote Sensing Imaging Point Clouds

This research introduces an innovative approach to address the vulnerability of deep learning-based point cloud target recognition systems to adversarial sample attacks. Instead of directly tampering with the spatial location information of the point cloud data, the approach focuses on modifying specific attribute information. The modified point cloud data is then associated with a simulated physical environment for comprehensive attack testing. The experimental findings demonstrate the aggressive nature of the generated counter-samples achieved through signal amplitude modulation, effectively deceiving deep learning-based target recognition systems. The experimental results highlight the effectiveness of the adversarial object samples generated by modifying the signal amplitude of the point cloud data, showcasing their strong misguiding capabilities towards the deep learning-based target recognition algorithm. This approach ensures good stealthiness and practicality, making it a viable attack method applicable in physical scenarios.