Abstract:To verify the vulnerability of LiDAR-based perception algorithms in autonomous driving when facing attacks, a black-box attack model is proposed. This model uses the PC2-PU point clouds upsampling algorithm in combination with random downsampling to control the number of attack points. This attack can reduce the occurrence of biases in the results caused by inherent flaws in the dataset. The proposed attack model includes three main factors: the number, location, and height of attack points. Experiments are conducted to verify the impact of these factors on the attack outcomes. Four dif-ferent types of perception models are used in the experiments to demonstrate the effectiveness of the proposed attack. The results show that even with only 20 attack points, the attack success rate exceeds 90% in some interval positions. As the number of attack points increases, the attack success rate continues to improve at different positions. The impact of this attack on the decision-making layer is also tested on the Baidu Apollo platform.